Upload 6 files

aworld/agents/README-multi-agent.md

@@ -0,0 +1,255 @@
+# Multi-agent
+
+```python
+from aworld.agents.llm_agent import Agent
+from aworld.config.conf import AgentConfig
+from aworld.core.agent.swarm import Swarm, GraphBuildType
+
+agent_conf = AgentConfig(...)
+```
+
+## Builder
+Builder represents the way topology is constructed, which is related to runtime execution. 
+Topology is the definition of structure. For the same topology structure, different builders 
+will produce execution processes and different results.
+
+```python
+"""
+Topology:
+ ┌─────A─────┐     
+ B     |     C
+       D 
+"""
+A = Agent(name="A", conf=agent_conf)
+B = Agent(name="B", conf=agent_conf)
+C = Agent(name="C", conf=agent_conf)
+D = Agent(name="D", conf=agent_conf)
+```
+
+### Workflow
+Workflow is a special topological structure that can be executed deterministically, all nodes in the swarm 
+will be executed. And the starting and ending nodes are **unique** and **indispensable**.
+
+Define:
+```python
+# default is workflow
+Swarm((A, B), (A, C), (A, D))
+or
+Swarm(A, [B, C, D])
+```
+The example means A is the start node, and the merge of B, C, and D is the end node.
+
+### Handoff
+Handoff using pure AI to drive the flow of the entire topology diagram, one agent's decision hands off 
+control to another. Agents as tools, depending on the defined pairs of agents.
+
+Define:
+```python
+Swarm((A, B), (A, C), (A, D), build_type=GraphBuildType.HANDOFF)
+or
+HandoffSwarm((A, B), (A, C), (A, D))
+```
+**NOTE**: Handoff supported tuple of paired agents forms only.
+
+### Team
+Team requires a leadership agent, and other agents follow its command. 
+Team is a special case of handoff, which is the leader-follower mode.
+
+Define:
+```python
+Swarm((A, B), (A, C), (A, D), build_type=GraphBuildType.TEAM)
+or
+TeamSwarm(A, B, C, D)
+or
+Swarm(B, C, D, root_agent=A, build_type=GraphBuildType.TEAM)
+```
+The root_agent or first agent A is the leader; other agents interact with the leader A.
+
+
+## Topology
+The topology structure of multi-agent is represented by Swarm, Swarm's topology is built based on 
+various single agents，can use the topology type and build type Swarm to represent different structural types.
+
+### Star
+Each agent communicates with a single supervisor agent, also known as star topology, 
+a special structure of tree topology, also referred to as a team topology in **Aworld**.
+
+A plan agent with other executing agents is a typical example.
+```python
+"""
+Star topology:
+ ┌───── plan ───┐     
+exec1         exec2
+"""
+plan = Agent(name="plan", conf=agent_conf)
+exec1 = Agent(name="exec1", conf=agent_conf)
+exec2 = Agent(name="exec2", conf=agent_conf)
+```
+
+We have two ways to construct this topology structure.
+```python
+swarm = Swarm((plan, exec1), (plan, exec2))
+```
+or use handoffs mechanism:
+```python
+plan = Agent(name="plan", conf=agent_conf, agent_names=['exec1', 'exec2'])
+swarm = Swarm(plan, register_agents=[exec1, exec2])
+```
+or use team mechanism:
+```python
+# The order of the plan agent is the first.
+swarm = TeamSwarm(plan, exec1, exec2,
+                 build_type=GraphBuildType.TEAM)
+```
+
+Note: 
+- Whether to execute exec1 or exec2 is decided by LLM.
+- If you want to execute all defined nodes with certainty, you need to use the `workflow` pattern.
+Like this will execute all the defined nodes:
+```python
+swarm = Swarm(plan, [exec1, exec2])
+```
+- If it is necessary to execute exec1, whether to execute exec2 depends on LLM, you can define it as:
+```python
+plan = Agent(name="plan", conf=agent_conf, agent_names=['exec1', 'exec2'])
+swarm = Swarm((plan, exec1), register_agents=[exec2])
+```
+That means that **GraphBuildType.WORKFLOW** is set, all nodes within the swarm will be executed.
+
+### Tree
+This is a generalization of the star topology and allows for more complex control flows.
+
+#### Hierarchical
+```python
+"""
+Hierarchical topology:
+            ┌─────────── root ───────────┐
+  ┌───── parent1 ───┐        ┌─────── parent2 ───────┐
+leaf1_1         leaf1_2    leaf1_1                leaf2_2
+"""
+
+root = Agent(name="root", conf=agent_conf)
+parent1 = Agent(name="parent1", conf=agent_conf)
+parent2 = Agent(name="parent2", conf=agent_conf)
+leaf1_1 = Agent(name="leaf1_1", conf=agent_conf)
+leaf1_2 = Agent(name="leaf1_2", conf=agent_conf)
+leaf2_1 = Agent(name="leaf2_1", conf=agent_conf)
+leaf2_2 = Agent(name="leaf2_2", conf=agent_conf)
+```
+
+```python
+swarm = Swarm((root, parent1), (root, parent2), 
+              (parent1, leaf1_1), (parent1, leaf1_2), 
+              (parent2, leaf2_1), (parent2, leaf2_2),
+              build_type=GraphBuildType.HANDOFF)
+```
+or use agent handoff:
+```python
+root = Agent(name="root", conf=agent_conf, agent_names=['parent1', 'parent2'])
+parent1 = Agent(name="parent1", conf=agent_conf, agent_names=['leaf1_1', 'leaf1_2'])
+parent2 = Agent(name="parent2", conf=agent_conf, agent_names=['leaf2_1', 'leaf2_2'])
+
+swarm = HandoffSwarm((root, parent1), (root, parent2),
+                     register_agents=[leaf1_1, leaf1_2, leaf2_1, leaf2_2])
+```
+
+#### Map-reduce
+If the topology structure becomes further complex:
+```
+            ┌─────────── root ───────────┐
+  ┌───── parent1 ───┐        ┌────── parent2 ──────┐
+leaf1_1         leaf1_2   leaf1_1               leaf2_2
+  └─────result1─────┘        └───────result2───────┘   
+            └───────────final───────────┘ 
+```
+We define it as **Map-reduce** topology, equivalent to workflow in terms of execution mode. 
+
+Build in this way:
+
+```python
+result1 = Agent(name="result1", conf=agent_conf)
+result2 = Agent(name="result2", conf=agent_conf)
+final = Agent(name="final", conf=agent_conf)
+
+swarm = Swarm(
+    (root, [parent1, parent2]), 
+    (parent1, [leaf1_1, leaf1_2]), 
+    (parent2, [leaf2_1, leaf2_2]),
+    ([leaf1_1, leaf1_2], result1), 
+    ([leaf2_1, leaf2_2], result2),
+    ([result1, result2], final)
+)
+```
+Assuming there is a cycle final -> root in the topology, define it as:
+```python
+final = LoopableAgent(name="final", 
+                      conf=agent_conf, 
+                      max_run_times=5, 
+                      loop_point=root.name(), 
+                      stop_func=...)
+```
+`stop_func` is a function that determines whether to terminate prematurely.
+
+
+### Mesh
+Divided into a fully meshed topology and a partially meshed topology. 
+Fully meshed topology means that each agent can communicate with every other agent, 
+any agent can decide which other agent to call next.
+
+```python
+"""
+Fully Meshed topology:
+    ┌─────────── A ──────────┐
+    B ───────────|────────── C 
+    └─────────── D  ─────────┘
+"""
+A = Agent(name="A", conf=agent_conf)
+B = Agent(name="B", conf=agent_conf)
+C = Agent(name="C", conf=agent_conf)
+D = Agent(name="D", conf=agent_conf)
+```
+
+Network topology need to use the `handoffs` mechanism:
+```python
+swarm = HandoffsSwarm((A, B), (B, A),
+                      (A, C), (C, A),
+                      (A, D), (D, A),
+                      (B, C), (C, B),
+                      (B, D), (D, B),
+                      (C, D), (D, C))
+```
+If a few pairs are removed, it becomes a partially meshed topology.
+
+### Ring
+A ring topology structure is a closed loop formed by nodes.
+
+```python
+"""
+Ring topology:
+    ┌───────────> A >──────────┐
+    B                          C 
+    └───────────< D  <─────────┘
+"""
+A = Agent(name="A", conf=agent_conf)
+B = Agent(name="B", conf=agent_conf)
+C = Agent(name="C", conf=agent_conf)
+D = Agent(name="D", conf=agent_conf)
+```
+
+
+```python
+swarm = Swarm((A, C), (C, D), (D, B), (B, A))
+```
+**Note:**
+- This defined loop can only be executed once.
+- If you want to execute multiple times, need to define it as:
+
+```python
+B = LoopableAgent(name="B", max_run_times=5, stop_func=...)
+swarm = Swarm((A, C), (C, D), (D, B))
+```
+### hybrid
+A generalization of topology, supporting an arbitrary combination of topologies, internally capable of 
+loops, parallel, serial dependencies, and groups.
+
+## Execution

aworld/agents/README.md

@@ -0,0 +1,107 @@
+# AI Agents
+
+Intelligent agents that control devices or tools in env using AI models or policy.
+
+![Agent Architecture](../../readme_assets/framework_agent.png)
+
+Most of the time, we directly use existing tools to build different types of agents that use LLM, 
+using frameworks makes it easy to write various agents.
+
+Detailed steps for building an agent:
+1. Define your `Agent`
+2. Write prompt used to the agent, also choose not to set it.
+3. Run it.
+    
+We provide a complete and simple example for writing an agent and multi-agent:
+
+```python
+from aworld.config.conf import AgentConfig
+from aworld.agents.llm_agent import Agent
+
+prompt = """
+Please act as a search agent, constructing appropriate keywords and searach terms, using search toolkit to collect relevant information, including urls, webpage snapshots, etc.
+Here are some tips that help you perform web search:
+- Never add too many keywords in your search query! Some detailed results need to perform browser interaction to get, not using search toolkit.
+- If the question is complex, search results typically do not provide precise answers. It is not likely to find the answer directly using search toolkit only, the search query should be concise and focuses on finding official sources rather than direct answers.
+  For example, as for the question "What is the maximum length in meters of #9 in the first National Geographic short on YouTube that was ever released according to the Monterey Bay Aquarium website?", your first search term must be coarse-grained like "National Geographic YouTube" to find the youtube website first, and then try other fine-grained search terms step-by-step to find more urls.
+- The results you return do not have to directly answer the original question, you only need to collect relevant information.
+
+Here are the question: {task}
+
+Please perform web search and return the listed search result, including urls and necessary webpage snapshots, introductions, etc.
+Your output should be like the followings (at most 3 relevant pages from coa):
+[
+    {{
+        "url": [URL],
+        "information": [INFORMATION OR CONTENT]
+    }},
+    ...
+]
+"""
+
+# Step1
+agent_config = AgentConfig(
+    llm_provider="openai",
+    llm_model_name="gpt-4o",
+    llm_temperature=1,
+    # need to set llm_api_key for use LLM
+    llm_api_key=""
+)
+
+search = Agent(
+    conf=agent_config,
+    name="search_agent",
+    system_prompt="You are a helpful search agent.",
+    # used to opt the result, also choose not to set it
+    agent_prompt=prompt,
+    tool_names=["search_api"]
+)
+
+```
+
+It can also quickly develop multi-agent based on the framework.
+
+On the basis of the above agent(SearchAgent), we provide a multi-agent example:
+
+```python
+from aworld.agents.llm_agent import Agent
+
+summary_prompt = """
+Summarize the following text in one clear and concise paragraph, capturing the key ideas without missing critical points. 
+Ensure the summary is easy to understand and avoids excessive detail.
+
+Here are the content: 
+{task}
+"""
+
+summary = Agent(
+    conf=agent_config,
+    name="summary_agent",
+    system_prompt="You are a helpful general summary agent.",
+    # used to opt the result, also choose not to set it
+    agent_prompt=summary_prompt
+)
+```
+
+You can run single-agent or multi-agent through Swarm.
+NOTE: Need to set some environment variables first! Effective GOOGLE_API_KEY, GOOGLE_ENGINE_ID, OPENAI_API_KEY and OPENAI_ENDPOINT.
+
+```python
+
+from aworld.core.agent.swarm import Swarm
+from aworld.runner import Runners
+
+if __name__ == '__main__':
+    task = "search 1+1=?"
+    # build topology graph, the correct order is necessary
+    swarm = Swarm(search, summary, max_steps=1)
+
+    prefix = ""
+    # can special search google, wiki, duck go, or baidu. such as:
+    # prefix = "search wiki: "
+    res = Runners.sync_run(
+        input=prefix + """What is an agent.""",
+        swarm=swarm
+    )
+```
+You can view search example [code](../../examples/search).

aworld/agents/llm_agent.py

@@ -0,0 +1,1011 @@
+# coding: utf-8
+# Copyright (c) 2025 inclusionAI.
+import abc
+import json
+import time
+import traceback
+import uuid
+from collections import OrderedDict
+from typing import AsyncGenerator, Dict, Any, List, Union, Callable
+
+import aworld.trace as trace
+from aworld.config import ToolConfig
+from aworld.config.conf import AgentConfig, ConfigDict, ContextRuleConfig, ModelConfig, OptimizationConfig, \
+    LlmCompressionConfig
+from aworld.core.agent.agent_desc import get_agent_desc
+from aworld.core.agent.base import BaseAgent, AgentResult, is_agent_by_name, is_agent
+from aworld.core.common import Observation, ActionModel
+from aworld.core.context.base import AgentContext
+from aworld.core.context.base import Context
+from aworld.core.context.processor.prompt_processor import PromptProcessor
+from aworld.core.event import eventbus
+from aworld.core.event.base import Message, ToolMessage, Constants, AgentMessage
+from aworld.core.tool.base import ToolFactory, AsyncTool, Tool
+from aworld.core.memory import MemoryItem, MemoryConfig
+from aworld.core.tool.tool_desc import get_tool_desc
+from aworld.logs.util import logger, color_log, Color, trace_logger
+from aworld.mcp_client.utils import sandbox_mcp_tool_desc_transform
+from aworld.memory.main import MemoryFactory
+from aworld.models.llm import get_llm_model, call_llm_model, acall_llm_model, acall_llm_model_stream
+from aworld.models.model_response import ModelResponse, ToolCall
+from aworld.models.utils import tool_desc_transform, agent_desc_transform
+from aworld.output import Outputs
+from aworld.output.base import StepOutput, MessageOutput
+from aworld.runners.hook.hook_factory import HookFactory
+from aworld.runners.hook.hooks import HookPoint
+from aworld.utils.common import sync_exec, nest_dict_counter
+
+
+class Agent(BaseAgent[Observation, List[ActionModel]]):
+    """Basic agent for unified protocol within the framework."""
+
+    def __init__(self,
+                 conf: Union[Dict[str, Any], ConfigDict, AgentConfig],
+                 resp_parse_func: Callable[..., Any] = None,
+                 **kwargs):
+        """A api class implementation of agent, using the `Observation` and `List[ActionModel]` protocols.
+
+        Args:
+            conf: Agent config, supported AgentConfig, ConfigDict or dict.
+            resp_parse_func: Response parse function for the agent standard output, transform llm response.
+        """
+        super(Agent, self).__init__(conf, **kwargs)
+        conf = self.conf
+        self.model_name = conf.llm_config.llm_model_name if conf.llm_config.llm_model_name else conf.llm_model_name
+        self._llm = None
+        self.memory = MemoryFactory.from_config(MemoryConfig(provider="inmemory"))
+        self.system_prompt: str = kwargs.pop("system_prompt") if kwargs.get("system_prompt") else conf.system_prompt
+        self.agent_prompt: str = kwargs.get("agent_prompt") if kwargs.get("agent_prompt") else conf.agent_prompt
+
+        self.event_driven = kwargs.pop('event_driven', conf.get('event_driven', False))
+        self.handler: Callable[..., Any] = kwargs.get('handler')
+
+        self.need_reset = kwargs.get('need_reset') if kwargs.get('need_reset') else conf.need_reset
+        # whether to keep contextual information, False means keep, True means reset in every step by the agent call
+        self.step_reset = kwargs.get('step_reset') if kwargs.get('step_reset') else True
+        # tool_name: [tool_action1, tool_action2, ...]
+        self.black_tool_actions: Dict[str, List[str]] = kwargs.get("black_tool_actions") if kwargs.get(
+            "black_tool_actions") else conf.get('black_tool_actions', {})
+        self.resp_parse_func = resp_parse_func if resp_parse_func else self.response_parse
+        self.history_messages = kwargs.get("history_messages") if kwargs.get("history_messages") else 100
+        self.use_tools_in_prompt = kwargs.get('use_tools_in_prompt', conf.use_tools_in_prompt)
+        self.context_rule = kwargs.get("context_rule") if kwargs.get("context_rule") else conf.context_rule
+        self.tools_instances = {}
+        self.tools_conf = {}
+
+    def reset(self, options: Dict[str, Any]):
+        super().reset(options)
+        self.memory = MemoryFactory.from_config(
+            MemoryConfig(provider=options.pop("memory_store") if options.get("memory_store") else "inmemory"))
+
+    def set_tools_instances(self, tools, tools_conf):
+        self.tools_instances = tools
+        self.tools_conf = tools_conf
+
+    @property
+    def llm(self):
+        # lazy
+        if self._llm is None:
+            llm_config = self.conf.llm_config or None
+            conf = llm_config if llm_config and (
+                    llm_config.llm_provider or llm_config.llm_base_url or llm_config.llm_api_key or llm_config.llm_model_name) else self.conf
+            self._llm = get_llm_model(conf)
+        return self._llm
+
+    def _env_tool(self):
+        """Description of agent as tool."""
+        return tool_desc_transform(get_tool_desc(),
+                                   tools=self.tool_names if self.tool_names else [],
+                                   black_tool_actions=self.black_tool_actions)
+
+    def _handoffs_agent_as_tool(self):
+        """Description of agent as tool."""
+        return agent_desc_transform(get_agent_desc(),
+                                    agents=self.handoffs if self.handoffs else [])
+
+    def _mcp_is_tool(self):
+        """Description of mcp servers are tools."""
+        try:
+            return sync_exec(sandbox_mcp_tool_desc_transform, self.mcp_servers, self.mcp_config)
+        except Exception as e:
+            logger.error(f"mcp_is_tool error: {traceback.format_exc()}")
+            return []
+
+    def desc_transform(self):
+        """Transform of descriptions of supported tools, agents, and MCP servers in the framework to support function calls of LLM."""
+
+        # Stateless tool
+        self.tools = self._env_tool()
+        # Agents as tool
+        self.tools.extend(self._handoffs_agent_as_tool())
+        # MCP servers are tools
+        self.tools.extend(self._mcp_is_tool())
+        # load to context
+        self.agent_context.set_tools(self.tools)
+        return self.tools
+
+    async def async_desc_transform(self):
+        """Transform of descriptions of supported tools, agents, and MCP servers in the framework to support function calls of LLM."""
+
+        # Stateless tool
+        self.tools = self._env_tool()
+        # Agents as tool
+        self.tools.extend(self._handoffs_agent_as_tool())
+        # MCP servers are tools
+        # todo sandbox
+        if self.sandbox:
+            sand_box = self.sandbox
+            mcp_tools = await sand_box.mcpservers.list_tools()
+            self.tools.extend(mcp_tools)
+        else:
+            self.tools.extend(await sandbox_mcp_tool_desc_transform(self.mcp_servers, self.mcp_config))
+        # load to agent context
+        self.agent_context.set_tools(self.tools)
+
+    def _messages_transform(
+            self,
+            observation: Observation,
+    ):
+        agent_prompt = self.agent_context.agent_prompt
+        sys_prompt = self.agent_context.sys_prompt
+        messages = []
+        if sys_prompt:
+            messages.append(
+                {'role': 'system', 'content': sys_prompt if not self.use_tools_in_prompt else sys_prompt.format(
+                    tool_list=self.tools)})
+
+        content = observation.content
+        if agent_prompt and '{task}' in agent_prompt:
+            content = agent_prompt.format(task=observation.content)
+
+        cur_msg = {'role': 'user', 'content': content}
+        # query from memory,
+        # histories = self.memory.get_last_n(self.history_messages, filter={"session_id": self.context.session_id})
+        histories = self.memory.get_last_n(self.history_messages)
+        messages.extend(histories)
+
+        action_results = observation.action_result
+        if action_results:
+            for action_result in action_results:
+                cur_msg['role'] = 'tool'
+                cur_msg['tool_call_id'] = action_result.tool_id
+
+        agent_info = self.context.context_info.get(self.id())
+        if (self.use_tools_in_prompt and "is_use_tool_prompt" in agent_info and "tool_calls"
+                in agent_info and agent_prompt):
+            cur_msg['content'] = agent_prompt.format(action_list=agent_info["tool_calls"],
+                                                     result=content)
+
+        if observation.images:
+            urls = [{'type': 'text', 'text': content}]
+            for image_url in observation.images:
+                urls.append({'type': 'image_url', 'image_url': {"url": image_url}})
+
+            cur_msg['content'] = urls
+        messages.append(cur_msg)
+
+        # truncate and other process
+        try:
+            messages = self._process_messages(messages=messages, agent_context=self.agent_context, context=self.context)
+        except Exception as e:
+            logger.warning(f"Failed to process messages in _messages_transform: {e}")
+            logger.debug(f"Process messages error details: {traceback.format_exc()}")
+        self.agent_context.update_messages(messages)
+        return messages
+
+    def messages_transform(self,
+                           content: str,
+                           image_urls: List[str] = None,
+                           **kwargs):
+        """Transform the original content to LLM messages of native format.
+
+        Args:
+            content: User content.
+            image_urls: List of images encoded using base64.
+            sys_prompt: Agent system prompt.
+            max_step: The maximum list length obtained from memory.
+        Returns:
+            Message list for LLM.
+        """
+        sys_prompt = self.agent_context.system_prompt
+        agent_prompt = self.agent_context.agent_prompt
+        messages = []
+        if sys_prompt:
+            messages.append(
+                {'role': 'system', 'content': sys_prompt if not self.use_tools_in_prompt else sys_prompt.format(
+                    tool_list=self.tools)})
+
+        histories = self.memory.get_last_n(self.history_messages)
+        user_content = content
+        if not histories and agent_prompt and '{task}' in agent_prompt:
+            user_content = agent_prompt.format(task=content)
+
+        cur_msg = {'role': 'user', 'content': user_content}
+        # query from memory,
+        # histories = self.memory.get_last_n(self.history_messages, filter={"session_id": self.context.session_id})
+
+        if histories:
+            # default use the first tool call
+            for history in histories:
+                if not self.use_tools_in_prompt and "tool_calls" in history.metadata and history.metadata['tool_calls']:
+                    messages.append({'role': history.metadata['role'], 'content': history.content,
+                                     'tool_calls': [history.metadata["tool_calls"][0]]})
+                else:
+                    messages.append({'role': history.metadata['role'], 'content': history.content,
+                                     "tool_call_id": history.metadata.get("tool_call_id")})
+
+            if not self.use_tools_in_prompt and "tool_calls" in histories[-1].metadata and histories[-1].metadata[
+                'tool_calls']:
+                tool_id = histories[-1].metadata["tool_calls"][0].id
+                if tool_id:
+                    cur_msg['role'] = 'tool'
+                    cur_msg['tool_call_id'] = tool_id
+            if self.use_tools_in_prompt and "is_use_tool_prompt" in histories[-1].metadata and "tool_calls" in \
+                    histories[-1].metadata and agent_prompt:
+                cur_msg['content'] = agent_prompt.format(action_list=histories[-1].metadata["tool_calls"],
+                                                         result=content)
+
+        if image_urls:
+            urls = [{'type': 'text', 'text': content}]
+            for image_url in image_urls:
+                urls.append({'type': 'image_url', 'image_url': {"url": image_url}})
+
+            cur_msg['content'] = urls
+        messages.append(cur_msg)
+
+        # truncate and other process
+        try:
+            messages = self._process_messages(messages=messages, agent_context=self.agent_context, context=self.context)
+        except Exception as e:
+            logger.warning(f"Failed to process messages in messages_transform: {e}")
+            logger.debug(f"Process messages error details: {traceback.format_exc()}")
+        self.agent_context.set_messages(messages)
+        return messages
+
+    def use_tool_list(self, resp: ModelResponse) -> List[Dict[str, Any]]:
+        tool_list = []
+        try:
+            if resp and hasattr(resp, 'content') and resp.content:
+                content = resp.content.strip()
+            else:
+                return tool_list
+            content = content.replace('\n', '').replace('\r', '')
+            response_json = json.loads(content)
+            if "use_tool_list" in response_json:
+                use_tool_list = response_json["use_tool_list"]
+                if use_tool_list:
+                    for use_tool in use_tool_list:
+                        tool_name = use_tool["tool"]
+                        arguments = use_tool["arguments"]
+                        if tool_name and arguments:
+                            tool_list.append(use_tool)
+
+            return tool_list
+        except Exception as e:
+            logger.debug(f"tool_parse error, content: {resp.content}, \nerror msg: {traceback.format_exc()}")
+            return tool_list
+
+    def response_parse(self, resp: ModelResponse) -> AgentResult:
+        """Default parse response by LLM."""
+        results = []
+        if not resp:
+            logger.warning("LLM no valid response!")
+            return AgentResult(actions=[], current_state=None)
+
+        use_tool_list = self.use_tool_list(resp)
+        is_call_tool = False
+        content = '' if resp.content is None else resp.content
+        if resp.tool_calls:
+            is_call_tool = True
+            for tool_call in resp.tool_calls:
+                full_name: str = tool_call.function.name
+                if not full_name:
+                    logger.warning("tool call response no tool name.")
+                    continue
+                try:
+                    params = json.loads(tool_call.function.arguments)
+                except:
+                    logger.warning(f"{tool_call.function.arguments} parse to json fail.")
+                    params = {}
+                # format in framework
+                names = full_name.split("__")
+                tool_name = names[0]
+                if is_agent_by_name(tool_name):
+                    param_info = params.get('content', "") + ' ' + params.get('info', '')
+                    results.append(ActionModel(tool_name=tool_name,
+                                               tool_id=tool_call.id,
+                                               agent_name=self.id(),
+                                               params=params,
+                                               policy_info=content + param_info))
+                else:
+                    action_name = '__'.join(names[1:]) if len(names) > 1 else ''
+                    results.append(ActionModel(tool_name=tool_name,
+                                               tool_id=tool_call.id,
+                                               action_name=action_name,
+                                               agent_name=self.id(),
+                                               params=params,
+                                               policy_info=content))
+        elif use_tool_list and len(use_tool_list) > 0:
+            is_call_tool = True
+            for use_tool in use_tool_list:
+                full_name = use_tool["tool"]
+                if not full_name:
+                    logger.warning("tool call response no tool name.")
+                    continue
+                params = use_tool["arguments"]
+                if not params:
+                    logger.warning("tool call response no tool params.")
+                    continue
+                names = full_name.split("__")
+                tool_name = names[0]
+                if is_agent_by_name(tool_name):
+                    param_info = params.get('content', "") + ' ' + params.get('info', '')
+                    results.append(ActionModel(tool_name=tool_name,
+                                               tool_id=use_tool.get('id'),
+                                               agent_name=self.id(),
+                                               params=params,
+                                               policy_info=content + param_info))
+                else:
+                    action_name = '__'.join(names[1:]) if len(names) > 1 else ''
+                    results.append(ActionModel(tool_name=tool_name,
+                                               tool_id=use_tool.get('id'),
+                                               action_name=action_name,
+                                               agent_name=self.id(),
+                                               params=params,
+                                               policy_info=content))
+        else:
+            if content:
+                content = content.replace("```json", "").replace("```", "")
+            # no tool call, agent name is itself.
+            results.append(ActionModel(agent_name=self.id(), policy_info=content))
+        return AgentResult(actions=results, current_state=None, is_call_tool=is_call_tool)
+
+    def _log_messages(self, messages: List[Dict[str, Any]]) -> None:
+        """Log the sequence of messages for debugging purposes"""
+        logger.info(f"[agent] Invoking LLM with {len(messages)} messages:")
+        for i, msg in enumerate(messages):
+            prefix = msg.get('role')
+            logger.info(f"[agent] Message {i + 1}: {prefix} ===================================")
+            if isinstance(msg['content'], list):
+                for item in msg['content']:
+                    if item.get('type') == 'text':
+                        logger.info(f"[agent] Text content: {item.get('text')}")
+                    elif item.get('type') == 'image_url':
+                        image_url = item.get('image_url', {}).get('url', '')
+                        if image_url.startswith('data:image'):
+                            logger.info(f"[agent] Image: [Base64 image data]")
+                        else:
+                            logger.info(f"[agent] Image URL: {image_url[:30]}...")
+            else:
+                content = str(msg['content'])
+                chunk_size = 500
+                for j in range(0, len(content), chunk_size):
+                    chunk = content[j:j + chunk_size]
+                    if j == 0:
+                        logger.info(f"[agent] Content: {chunk}")
+                    else:
+                        logger.info(f"[agent] Content (continued): {chunk}")
+
+            if 'tool_calls' in msg and msg['tool_calls']:
+                for tool_call in msg.get('tool_calls'):
+                    if isinstance(tool_call, dict):
+                        logger.info(f"[agent] Tool call: {tool_call.get('name')} - ID: {tool_call.get('id')}")
+                        args = str(tool_call.get('args', {}))[:1000]
+                        logger.info(f"[agent] Tool args: {args}...")
+                    elif isinstance(tool_call, ToolCall):
+                        logger.info(f"[agent] Tool call: {tool_call.function.name} - ID: {tool_call.id}")
+                        args = str(tool_call.function.arguments)[:1000]
+                        logger.info(f"[agent] Tool args: {args}...")
+
+    def _agent_result(self, actions: List[ActionModel], caller: str):
+        if not actions:
+            raise Exception(f'{self.id()} no action decision has been made.')
+
+        tools = OrderedDict()
+        agents = []
+        for action in actions:
+            if is_agent(action):
+                agents.append(action)
+            else:
+                if action.tool_name not in tools:
+                    tools[action.tool_name] = []
+                tools[action.tool_name].append(action)
+
+        _group_name = None
+        # agents and tools exist simultaneously, more than one agent/tool name
+        if (agents and tools) or len(agents) > 1 or len(tools) > 1:
+            _group_name = f"{self.id()}_{uuid.uuid1().hex}"
+
+        # complex processing
+        if _group_name:
+            logger.warning(f"more than one agent an tool causing confusion, will choose the first one. {agents}")
+            agents = [agents[0]] if agents else []
+            for _, v in tools.items():
+                actions = v
+                break
+
+        if agents:
+            return AgentMessage(payload=actions,
+                                caller=caller,
+                                sender=self.id(),
+                                receiver=actions[0].tool_name,
+                                session_id=self.context.session_id if self.context else "",
+                                headers={"context": self.context})
+        else:
+            return ToolMessage(payload=actions,
+                               caller=caller,
+                               sender=self.id(),
+                               receiver=actions[0].tool_name,
+                               session_id=self.context.session_id if self.context else "",
+                               headers={"context": self.context})
+
+    def post_run(self, policy_result: List[ActionModel], policy_input: Observation) -> Message:
+        return self._agent_result(
+            policy_result,
+            policy_input.from_agent_name if policy_input.from_agent_name else policy_input.observer
+        )
+
+    async def async_post_run(self, policy_result: List[ActionModel], policy_input: Observation) -> Message:
+        return self._agent_result(
+            policy_result,
+            policy_input.from_agent_name if policy_input.from_agent_name else policy_input.observer
+        )
+
+    def policy(self, observation: Observation, info: Dict[str, Any] = {}, **kwargs) -> List[ActionModel]:
+        """The strategy of an agent can be to decide which tools to use in the environment, or to delegate tasks to other agents.
+
+        Args:
+            observation: The state observed from tools in the environment.
+            info: Extended information is used to assist the agent to decide a policy.
+
+        Returns:
+            ActionModel sequence from agent policy
+        """
+        output = None
+        if kwargs.get("output") and isinstance(kwargs.get("output"), StepOutput):
+            output = kwargs["output"]
+
+        # Get current step information for trace recording
+        step = kwargs.get("step", 0)
+        exp_id = kwargs.get("exp_id", None)
+        source_span = trace.get_current_span()
+
+        if hasattr(observation, 'context') and observation.context:
+            self.task_histories = observation.context
+
+        try:
+            self._run_hooks_sync(self.context, HookPoint.PRE_LLM_CALL)
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+
+        self._finished = False
+        self.desc_transform()
+        images = observation.images if self.conf.use_vision else None
+        if self.conf.use_vision and not images and observation.image:
+            images = [observation.image]
+            observation.images = images
+        messages = self.messages_transform(content=observation.content,
+                                           image_urls=observation.images)
+
+        self._log_messages(messages)
+        self.memory.add(MemoryItem(
+            content=messages[-1]['content'],
+            metadata={
+                "role": messages[-1]['role'],
+                "agent_name": self.id(),
+                "tool_call_id": messages[-1].get("tool_call_id")
+            }
+        ))
+
+        llm_response = None
+        span_name = f"llm_call_{exp_id}"
+        serializable_messages = self._to_serializable(messages)
+        with trace.span(span_name) as llm_span:
+            llm_span.set_attributes({
+                "exp_id": exp_id,
+                "step": step,
+                "messages": json.dumps(serializable_messages, ensure_ascii=False)
+            })
+            if source_span:
+                source_span.set_attribute("messages", json.dumps(serializable_messages, ensure_ascii=False))
+
+            try:
+                llm_response = call_llm_model(
+                    self.llm,
+                    messages=messages,
+                    model=self.model_name,
+                    temperature=self.conf.llm_config.llm_temperature,
+                    tools=self.tools if not self.use_tools_in_prompt and self.tools else None
+                )
+
+                logger.info(f"Execute response: {llm_response.message}")
+            except Exception as e:
+                logger.warn(traceback.format_exc())
+                raise e
+            finally:
+                if llm_response:
+                    # update usage
+                    self.update_context_usage(used_context_length=llm_response.usage['total_tokens'])
+                    # update current step output
+                    self.update_llm_output(llm_response)
+
+                    use_tools = self.use_tool_list(llm_response)
+                    is_use_tool_prompt = len(use_tools) > 0
+                    if llm_response.error:
+                        logger.info(f"llm result error: {llm_response.error}")
+                    else:
+                        info = {
+                            "role": "assistant",
+                            "agent_name": self.id(),
+                            "tool_calls": llm_response.tool_calls if not self.use_tools_in_prompt else use_tools,
+                            "is_use_tool_prompt": is_use_tool_prompt if not self.use_tools_in_prompt else False
+                        }
+                        self.memory.add(MemoryItem(
+                            content=llm_response.content,
+                            metadata=info
+                        ))
+                        # rewrite
+                        self.context.context_info[self.id()] = info
+                else:
+                    logger.error(f"{self.id()} failed to get LLM response")
+                    raise RuntimeError(f"{self.id()} failed to get LLM response")
+
+        try:
+            self._run_hooks_sync(self.context, HookPoint.POST_LLM_CALL)
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+
+        agent_result = sync_exec(self.resp_parse_func, llm_response)
+        if not agent_result.is_call_tool:
+            self._finished = True
+
+        if output:
+            output.add_part(MessageOutput(source=llm_response, json_parse=False))
+            output.mark_finished()
+        return agent_result.actions
+
+    async def async_policy(self, observation: Observation, info: Dict[str, Any] = {}, **kwargs) -> List[ActionModel]:
+        """The strategy of an agent can be to decide which tools to use in the environment, or to delegate tasks to other agents.
+
+        Args:
+            observation: The state observed from tools in the environment.
+            info: Extended information is used to assist the agent to decide a policy.
+
+        Returns:
+            ActionModel sequence from agent policy
+        """
+        outputs = None
+        if kwargs.get("outputs") and isinstance(kwargs.get("outputs"), Outputs):
+            outputs = kwargs.get("outputs")
+
+        # Get current step information for trace recording
+        source_span = trace.get_current_span()
+
+        if hasattr(observation, 'context') and observation.context:
+            self.task_histories = observation.context
+
+        try:
+            events = []
+            async for event in self.run_hooks(self.context, HookPoint.PRE_LLM_CALL):
+                events.append(event)
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+
+        self._finished = False
+        messages = await self._prepare_llm_input(observation, info, **kwargs)
+
+        serializable_messages = self._to_serializable(messages)
+        llm_response = None
+        if source_span:
+            source_span.set_attribute("messages", json.dumps(serializable_messages, ensure_ascii=False))
+        try:
+            llm_response = await self._call_llm_model(observation, messages, info, **kwargs)
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+            raise e
+        finally:
+            if llm_response:
+                # update usage
+                self.update_context_usage(used_context_length=llm_response.usage['total_tokens'])
+                # update current step output
+                self.update_llm_output(llm_response)
+
+                use_tools = self.use_tool_list(llm_response)
+                is_use_tool_prompt = len(use_tools) > 0
+                if llm_response.error:
+                    logger.info(f"llm result error: {llm_response.error}")
+                else:
+                    self.memory.add(MemoryItem(
+                        content=llm_response.content,
+                        metadata={
+                            "role": "assistant",
+                            "agent_name": self.id(),
+                            "tool_calls": llm_response.tool_calls if not self.use_tools_in_prompt else use_tools,
+                            "is_use_tool_prompt": is_use_tool_prompt if not self.use_tools_in_prompt else False
+                        }
+                    ))
+            else:
+                logger.error(f"{self.id()} failed to get LLM response")
+                raise RuntimeError(f"{self.id()} failed to get LLM response")
+
+        try:
+            events = []
+            async for event in self.run_hooks(self.context, HookPoint.POST_LLM_CALL):
+                events.append(event)
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+
+        agent_result = sync_exec(self.resp_parse_func, llm_response)
+        if not agent_result.is_call_tool:
+            self._finished = True
+        return agent_result.actions
+
+    def _to_serializable(self, obj):
+        if isinstance(obj, dict):
+            return {k: self._to_serializable(v) for k, v in obj.items()}
+        elif isinstance(obj, list):
+            return [self._to_serializable(i) for i in obj]
+        elif hasattr(obj, "to_dict"):
+            return obj.to_dict()
+        elif hasattr(obj, "model_dump"):
+            return obj.model_dump()
+        elif hasattr(obj, "dict"):
+            return obj.dict()
+        else:
+            return obj
+
+    async def llm_and_tool_execution(self, observation: Observation, messages: List[Dict[str, str]] = [],
+                                     info: Dict[str, Any] = {}, **kwargs) -> List[ActionModel]:
+        """Perform combined LLM call and tool execution operations.
+
+        Args:
+            observation: The state observed from the environment
+            info: Extended information to assist the agent in decision-making
+            **kwargs: Other parameters
+
+        Returns:
+            ActionModel sequence. If a tool is executed, includes the tool execution result.
+        """
+        # Get current step information for trace recording
+        llm_response = await self._call_llm_model(observation, messages, info, **kwargs)
+        if llm_response:
+            use_tools = self.use_tool_list(llm_response)
+            is_use_tool_prompt = len(use_tools) > 0
+            if llm_response.error:
+                logger.info(f"llm result error: {llm_response.error}")
+            else:
+                self.memory.add(MemoryItem(
+                    content=llm_response.content,
+                    metadata={
+                        "role": "assistant",
+                        "agent_name": self.id(),
+                        "tool_calls": llm_response.tool_calls if not self.use_tools_in_prompt else use_tools,
+                        "is_use_tool_prompt": is_use_tool_prompt if not self.use_tools_in_prompt else False
+                    }
+                ))
+        else:
+            logger.error(f"{self.id()} failed to get LLM response")
+            raise RuntimeError(f"{self.id()} failed to get LLM response")
+
+        agent_result = sync_exec(self.resp_parse_func, llm_response)
+        if not agent_result.is_call_tool:
+            self._finished = True
+            return agent_result.actions
+        else:
+            result = await self._execute_tool(agent_result.actions)
+            return result
+
+    async def _prepare_llm_input(self, observation: Observation, info: Dict[str, Any] = {}, **kwargs):
+        """Prepare LLM input
+        Args:
+            observation: The state observed from the environment
+            info: Extended information to assist the agent in decision-making
+            **kwargs: Other parameters
+        """
+        await self.async_desc_transform()
+        images = observation.images if self.conf.use_vision else None
+        if self.conf.use_vision and not images and observation.image:
+            images = [observation.image]
+        messages = self.messages_transform(content=observation.content,
+                                           image_urls=images)
+
+        self._log_messages(messages)
+        self.memory.add(MemoryItem(
+            content=messages[-1]['content'],
+            metadata={
+                "role": messages[-1]['role'],
+                "agent_name": self.id(),
+                "tool_call_id": messages[-1].get("tool_call_id")
+            }
+        ))
+
+        return messages
+
+    def _process_messages(self, messages: List[Dict[str, Any]], agent_context: AgentContext = None,
+                          context: Context = None) -> Message:
+        origin_messages = messages
+        st = time.time()
+        with trace.span(f"llm_context_process", attributes={
+            "start_time": st
+        }) as compress_span:
+            if agent_context.context_rule is None:
+                logger.debug('debug|skip process_messages context_rule is None')
+                return messages
+            origin_len = compressed_len = len(str(messages))
+            origin_messages_count = truncated_messages_count = len(messages)
+            try:
+                prompt_processor = PromptProcessor(agent_context)
+                result = prompt_processor.process_messages(messages, context)
+                messages = result.processed_messages
+
+                compressed_len = len(str(messages))
+                truncated_messages_count = len(messages)
+                logger.debug(
+                    f'debug|llm_context_process|{origin_len}|{compressed_len}|{origin_messages_count}|{truncated_messages_count}|\n|{origin_messages}\n|{messages}')
+                return messages
+            finally:
+                compress_span.set_attributes({
+                    "end_time": time.time(),
+                    "duration": time.time() - st,
+                    # messages length
+                    "origin_messages_count": origin_messages_count,
+                    "truncated_messages_count": truncated_messages_count,
+                    "truncated_ratio": round(truncated_messages_count / origin_messages_count, 2),
+                    # token length
+                    "origin_len": origin_len,
+                    "compressed_len": compressed_len,
+                    "compress_ratio": round(compressed_len / origin_len, 2)
+                })
+
+    async def _call_llm_model(self, observation: Observation, messages: List[Dict[str, str]] = [],
+                              info: Dict[str, Any] = {}, **kwargs) -> ModelResponse:
+        """Perform LLM call
+        Args:
+            observation: The state observed from the environment
+            info: Extended information to assist the agent in decision-making
+            **kwargs: Other parameters
+        Returns:
+            LLM response
+        """
+        outputs = None
+        if kwargs.get("outputs") and isinstance(kwargs.get("outputs"), Outputs):
+            outputs = kwargs.get("outputs")
+        if not messages:
+            messages = await self._prepare_llm_input(observation, self.agent_context, **kwargs)
+
+        llm_response = None
+        source_span = trace.get_current_span()
+        serializable_messages = self._to_serializable(messages)
+
+        if source_span:
+            source_span.set_attribute("messages", json.dumps(serializable_messages, ensure_ascii=False))
+
+        try:
+            stream_mode = kwargs.get("stream", False)
+            if stream_mode:
+                llm_response = ModelResponse(id="", model="", content="", tool_calls=[])
+                resp_stream = acall_llm_model_stream(
+                    self.llm,
+                    messages=messages,
+                    model=self.model_name,
+                    temperature=self.conf.llm_config.llm_temperature,
+                    tools=self.tools if not self.use_tools_in_prompt and self.tools else None,
+                    stream=True
+                )
+
+                async def async_call_llm(resp_stream, json_parse=False):
+                    llm_resp = ModelResponse(id="", model="", content="", tool_calls=[])
+
+                    # Async streaming with acall_llm_model
+                    async def async_generator():
+                        async for chunk in resp_stream:
+                            if chunk.content:
+                                llm_resp.content += chunk.content
+                                yield chunk.content
+                            if chunk.tool_calls:
+                                llm_resp.tool_calls.extend(chunk.tool_calls)
+                            if chunk.error:
+                                llm_resp.error = chunk.error
+                            llm_resp.id = chunk.id
+                            llm_resp.model = chunk.model
+                            llm_resp.usage = nest_dict_counter(llm_resp.usage, chunk.usage)
+
+                    return MessageOutput(source=async_generator(), json_parse=json_parse), llm_resp
+
+                output, response = await async_call_llm(resp_stream)
+                llm_response = response
+
+                if eventbus is not None and resp_stream:
+                    output_message = Message(
+                        category=Constants.OUTPUT,
+                        payload=output,
+                        sender=self.id(),
+                        session_id=self.context.session_id if self.context else "",
+                        headers={"context": self.context}
+                    )
+                    await eventbus.publish(output_message)
+                elif not self.event_driven and outputs:
+                    outputs.add_output(output)
+
+            else:
+                llm_response = await acall_llm_model(
+                    self.llm,
+                    messages=messages,
+                    model=self.model_name,
+                    temperature=self.conf.llm_config.llm_temperature,
+                    tools=self.tools if not self.use_tools_in_prompt and self.tools else None,
+                    stream=kwargs.get("stream", False)
+                )
+                if eventbus is None:
+                    logger.warn("=============== eventbus is none ============")
+                if eventbus is not None and llm_response:
+                    await eventbus.publish(Message(
+                        category=Constants.OUTPUT,
+                        payload=llm_response,
+                        sender=self.id(),
+                        session_id=self.context.session_id if self.context else "",
+                        headers={"context": self.context}
+                    ))
+                elif not self.event_driven and outputs:
+                    outputs.add_output(MessageOutput(source=llm_response, json_parse=False))
+
+            logger.info(f"Execute response: {json.dumps(llm_response.to_dict(), ensure_ascii=False)}")
+
+
+        except Exception as e:
+            logger.warn(traceback.format_exc())
+            raise e
+        finally:
+            return llm_response
+
+    async def _execute_tool(self, actions: List[ActionModel]) -> Any:
+        """Execute tool calls
+
+        Args:
+            action: The action(s) to execute
+
+        Returns:
+            The result of tool execution
+        """
+        tool_actions = []
+        for act in actions:
+            if is_agent(act):
+                continue
+            else:
+                tool_actions.append(act)
+
+        msg = None
+        terminated = False
+        # group action by tool name
+        tool_mapping = dict()
+        reward = 0.0
+        # Directly use or use tools after creation.
+        for act in tool_actions:
+            if not self.tools_instances or (self.tools_instances and act.tool_name not in self.tools):
+                # Dynamically only use default config in module.
+                conf = self.tools_conf.get(act.tool_name)
+                if not conf:
+                    conf = ToolConfig(exit_on_failure=self.task.conf.get('exit_on_failure'))
+                tool = ToolFactory(act.tool_name, conf=conf, asyn=conf.use_async if conf else False)
+                if isinstance(tool, Tool):
+                    tool.reset()
+                elif isinstance(tool, AsyncTool):
+                    await tool.reset()
+                tool_mapping[act.tool_name] = []
+                self.tools_instances[act.tool_name] = tool
+            if act.tool_name not in tool_mapping:
+                tool_mapping[act.tool_name] = []
+            tool_mapping[act.tool_name].append(act)
+
+        observation = None
+
+        for tool_name, action in tool_mapping.items():
+            # Execute action using browser tool and unpack all return values
+            if isinstance(self.tools_instances[tool_name], Tool):
+                message = self.tools_instances[tool_name].step(action)
+            elif isinstance(self.tools_instances[tool_name], AsyncTool):
+                # todo sandbox
+                message = await self.tools_instances[tool_name].step(action, agent=self)
+            else:
+                logger.warning(f"Unsupported tool type: {self.tools_instances[tool_name]}")
+                continue
+
+            observation, reward, terminated, _, info = message.payload
+
+            # Check if there's an exception in info
+            if info.get("exception"):
+                color_log(f"Agent {self.id()} _execute_tool failed with exception: {info['exception']}",
+                          color=Color.red)
+                msg = f"Agent {self.id()} _execute_tool failed with exception: {info['exception']}"
+            logger.info(f"Agent {self.id()} _execute_tool finished by tool action: {action}.")
+            log_ob = Observation(content='' if observation.content is None else observation.content,
+                                 action_result=observation.action_result)
+            trace_logger.info(f"{tool_name} observation: {log_ob}", color=Color.green)
+            self.memory.add(MemoryItem(
+                content=observation.content,
+                metadata={
+                    "role": "tool",
+                    "agent_name": self.id(),
+                    "tool_call_id": action[0].tool_id
+                }
+            ))
+        return [ActionModel(agent_name=self.id(), policy_info=observation.content)]
+
+    def _init_context(self, context: Context):
+        super()._init_context(context)
+        # Generate default configuration when context_rule is empty
+        llm_config = self.conf.llm_config
+        context_rule = self.context_rule
+        if context_rule is None:
+            context_rule = ContextRuleConfig(
+                optimization_config=OptimizationConfig(
+                    enabled=True,
+                    max_token_budget_ratio=1.0
+                ),
+                llm_compression_config=LlmCompressionConfig(
+                    enabled=False  # Compression disabled by default
+                )
+            )
+        self.agent_context.set_model_config(llm_config)
+        self.agent_context.context_rule = context_rule
+        self.agent_context.system_prompt = self.system_prompt
+        self.agent_context.agent_prompt = self.agent_prompt
+        logger.debug(f'init_context llm_agent {self.name()} {self.agent_context} {self.conf} {self.context_rule}')
+
+    def update_system_prompt(self, system_prompt: str):
+        self.system_prompt = system_prompt
+        self.agent_context.system_prompt = system_prompt
+        logger.info(f"Agent {self.name()} system_prompt updated")
+
+    def update_agent_prompt(self, agent_prompt: str):
+        self.agent_prompt = agent_prompt
+        self.agent_context.agent_prompt = agent_prompt
+        logger.info(f"Agent {self.name()} agent_prompt updated")
+
+    def update_context_rule(self, context_rule: ContextRuleConfig):
+        self.agent_context.context_rule = context_rule
+        logger.info(f"Agent {self.name()} context_rule updated")
+
+    def update_context_usage(self, used_context_length: int = None, total_context_length: int = None):
+        self.agent_context.update_context_usage(used_context_length, total_context_length)
+        logger.debug(f"Agent {self.name()} context usage updated: {self.agent_context.context_usage}")
+
+    def update_llm_output(self, llm_response: ModelResponse):
+        self.agent_context.set_llm_output(llm_response)
+        logger.debug(f"Agent {self.name()} llm output updated: {self.agent_context.llm_output}")
+
+    async def run_hooks(self, context: Context, hook_point: str):
+        """Execute hooks asynchronously"""
+        from aworld.runners.hook.hook_factory import HookFactory
+        from aworld.core.event.base import Message
+
+        # Get all hooks for the specified hook point
+        all_hooks = HookFactory.hooks(hook_point)
+        hooks = all_hooks.get(hook_point, [])
+
+        for hook in hooks:
+            try:
+                # Create a temporary Message object to pass to the hook
+                message = Message(
+                    category="agent_hook",
+                    payload=None,
+                    sender=self.id(),
+                    session_id=context.session_id if hasattr(context, 'session_id') else None,
+                    headers={"context": self.context}
+                )
+
+                # Execute hook
+                msg = await hook.exec(message, context)
+                if msg:
+                    logger.debug(f"Hook {hook.point()} executed successfully")
+                    yield msg
+            except Exception as e:
+                logger.warning(f"Hook {hook.point()} execution failed: {traceback.format_exc()}")
+
+    def _run_hooks_sync(self, context: Context, hook_point: str):
+        """Execute hooks synchronously"""
+        # Use sync_exec to execute asynchronous hook logic
+        try:
+            sync_exec(self.run_hooks, context, hook_point)
+        except Exception as e:
+            logger.warn(f"Failed to execute hooks for {hook_point}: {traceback.format_exc()}")

aworld/agents/loop_llm_agent.py

@@ -0,0 +1,46 @@
+# coding: utf-8
+# Copyright (c) 2025 inclusionAI.
+from typing import Any, Callable
+
+from aworld.agents.llm_agent import Agent
+
+
+class LoopableAgent(Agent):
+    """Support for loop agents in the swarm.
+
+    The parameters of the extension function are the agent itself, which can obtain internal information of the agent.
+    `stop_func` function example:
+    >>> def stop(agent: LoopableAgent):
+    >>>     ...
+
+    `loop_point_finder` function example:
+    >>> def find(agent: LoopableAgent):
+    >>>     ...
+    """
+    max_run_times: int = 1
+    cur_run_times: int = 0
+    # The loop agent special the loop point (agent name)
+    loop_point: str = None
+    # Used to determine the loop point for multiple loops
+    loop_point_finder: Callable[..., Any] = None
+    # def stop(agent: LoopableAgent): ...
+    stop_func: Callable[..., Any] = None
+
+    @property
+    def goto(self):
+        """The next loop point is what the loop agent wants to reach."""
+        if self.loop_point_finder:
+            return self.loop_point_finder(self)
+        if self.loop_point:
+            return self.loop_point
+        return self.id()
+
+    @property
+    def finished(self) -> bool:
+        """Loop agent termination state detection, achieved loop count or termination condition."""
+        if self.cur_run_times >= self.max_run_times or (self.stop_func and self.stop_func(self)):
+            self._finished = True
+            return True
+
+        self._finished = False
+        return False

aworld/agents/parallel_llm_agent.py

@@ -0,0 +1,54 @@
+# coding: utf-8
+# Copyright (c) 2025 inclusionAI.
+from typing import List, Dict, Any, Callable
+
+from aworld.agents.llm_agent import Agent
+from aworld.config import RunConfig
+from aworld.core.common import Observation, ActionModel, Config
+
+
+class ParallelizableAgent(Agent):
+    """Support for parallel agents in the swarm.
+
+    The parameters of the extension function are the agent itself, which can obtain internal information of the agent.
+    `aggregate_func` function example:
+    >>> def agg(agent: ParallelizableAgent, res: Dict[str, List[ActionModel]]):
+    >>>     ...
+    """
+
+    def __init__(self,
+                 conf: Config,
+                 resp_parse_func: Callable[..., Any] = None,
+                 agents: List[Agent] = [],
+                 aggregate_func: Callable[..., Any] = None,
+                 **kwargs):
+        super().__init__(conf=conf, resp_parse_func=resp_parse_func, **kwargs)
+        self.agents = agents
+        # The function of aggregating the results of the parallel execution of agents.
+        self.aggregate_func = aggregate_func
+
+    async def async_policy(self, observation: Observation, info: Dict[str, Any] = {}, **kwargs) -> List[ActionModel]:
+        from aworld.core.task import Task
+        from aworld.runners.utils import choose_runners, execute_runner
+
+        tasks = []
+        if self.agents:
+            for agent in self.agents:
+                tasks.append(Task(input=observation, agent=agent, context=self.context))
+
+        if not tasks:
+            raise RuntimeError("no task need to run in parallelizable agent.")
+
+        runners = await choose_runners(tasks)
+        res = await execute_runner(runners, RunConfig(reuse_process=False))
+
+        if self.aggregate_func:
+            return self.aggregate_func(self, res)
+
+        results = []
+        for k, v in res.items():
+            results.append(ActionModel(agent_name=self.id(), policy_info=v.answer))
+        return results
+
+    def finished(self) -> bool:
+        return all([agent.finished for agent in self.agents])

aworld/agents/serial_llm_agent.py

@@ -0,0 +1,48 @@
+# coding: utf-8
+# Copyright (c) 2025 inclusionAI.
+from typing import List, Dict, Any, Callable
+
+from aworld.agents.llm_agent import Agent
+from aworld.core.common import Observation, ActionModel, Config
+from aworld.logs.util import logger
+
+
+class SerialableAgent(Agent):
+    """Support for serial execution of agents based on dependency relationships in the swarm."""
+
+    def __init__(self,
+                 conf: Config,
+                 resp_parse_func: Callable[..., Any] = None,
+                 agents: List[Agent] = [],
+                 **kwargs):
+        super().__init__(conf=conf, resp_parse_func=resp_parse_func, **kwargs)
+        self.agents = agents
+
+    async def async_policy(self, observation: Observation, info: Dict[str, Any] = {}, **kwargs) -> List[ActionModel]:
+        from aworld.config import RunConfig
+        from aworld.core.task import Task
+        from aworld.runners.utils import choose_runners, execute_runner
+
+        action = ActionModel(agent_name=self.id(), policy_info=observation.content)
+        for agent in self.agents:
+            task = Task(input=observation, agent=agent, context=self.context)
+            runners = await choose_runners([task])
+            res = await execute_runner(runners, RunConfig(reuse_process=False))
+            if res:
+                v = res.get(task.id)
+                action = ActionModel(agent_name=self.id(), policy_info=v.answer)
+                observation = self._action_to_observation(action, agent.name())
+            else:
+                raise Exception(f"{agent.id()} execute fail.")
+        return [action]
+
+    def _action_to_observation(self, policy: ActionModel, agent_name: str):
+        if not policy:
+            logger.warning("no agent policy, will use default error info.")
+            return Observation(content=f"{agent_name} no policy")
+
+        logger.debug(f"{policy.policy_info}")
+        return Observation(content=policy.policy_info)
+
+    def finished(self) -> bool:
+        return all([agent.finished for agent in self.agents])