#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import importlib
import inspect
import json
import os
import re
import tempfile
import textwrap
import time
import warnings
from abc import ABC, abstractmethod
from collections.abc import Callable, Generator
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
from logging import getLogger
from pathlib import Path
from typing import TYPE_CHECKING, Any, Literal, TypeAlias, TypedDict, Union
import jinja2
import yaml
from huggingface_hub import create_repo, metadata_update, snapshot_download, upload_folder
from jinja2 import StrictUndefined, Template
from rich.console import Group
from rich.live import Live
from rich.markdown import Markdown
from rich.panel import Panel
from rich.rule import Rule
from rich.text import Text
if TYPE_CHECKING:
import PIL.Image
from .agent_types import AgentAudio, AgentImage, handle_agent_output_types
from .default_tools import TOOL_MAPPING, FinalAnswerTool
from .local_python_executor import BASE_BUILTIN_MODULES, LocalPythonExecutor, PythonExecutor, fix_final_answer_code
from .memory import (
ActionStep,
AgentMemory,
FinalAnswerStep,
PlanningStep,
SystemPromptStep,
TaskStep,
Timing,
TokenUsage,
ToolCall,
)
from .models import (
CODEAGENT_RESPONSE_FORMAT,
ChatMessage,
ChatMessageStreamDelta,
ChatMessageToolCall,
MessageRole,
Model,
agglomerate_stream_deltas,
parse_json_if_needed,
)
from .monitoring import (
YELLOW_HEX,
AgentLogger,
LogLevel,
Monitor,
)
from .remote_executors import DockerExecutor, E2BExecutor
from .tools import Tool, validate_tool_arguments
from .utils import (
AGENT_GRADIO_APP_TEMPLATE,
AgentError,
AgentExecutionError,
AgentGenerationError,
AgentMaxStepsError,
AgentParsingError,
AgentToolCallError,
AgentToolExecutionError,
extract_code_from_text,
is_valid_name,
make_init_file,
parse_code_blobs,
truncate_content,
)
logger = getLogger(__name__)
def get_variable_names(self, template: str) -> set[str]:
pattern = re.compile(r"\{\{([^{}]+)\}\}")
return {match.group(1).strip() for match in pattern.finditer(template)}
def populate_template(template: str, variables: dict[str, Any]) -> str:
compiled_template = Template(template, undefined=StrictUndefined)
try:
return compiled_template.render(**variables)
except Exception as e:
raise Exception(f"Error during jinja template rendering: {type(e).__name__}: {e}")
@dataclass
class ActionOutput:
output: Any
is_final_answer: bool
@dataclass
class ToolOutput:
id: str
output: Any
is_final_answer: bool
observation: str
tool_call: ToolCall
class PlanningPromptTemplate(TypedDict):
"""
Prompt templates for the planning step.
Args:
plan (`str`): Initial plan prompt.
update_plan_pre_messages (`str`): Update plan pre-messages prompt.
update_plan_post_messages (`str`): Update plan post-messages prompt.
"""
initial_plan: str
update_plan_pre_messages: str
update_plan_post_messages: str
class ManagedAgentPromptTemplate(TypedDict):
"""
Prompt templates for the managed agent.
Args:
task (`str`): Task prompt.
report (`str`): Report prompt.
"""
task: str
report: str
class FinalAnswerPromptTemplate(TypedDict):
"""
Prompt templates for the final answer.
Args:
pre_messages (`str`): Pre-messages prompt.
post_messages (`str`): Post-messages prompt.
"""
pre_messages: str
post_messages: str
class PromptTemplates(TypedDict):
"""
Prompt templates for the agent.
Args:
system_prompt (`str`): System prompt.
planning ([`~agents.PlanningPromptTemplate`]): Planning prompt templates.
managed_agent ([`~agents.ManagedAgentPromptTemplate`]): Managed agent prompt templates.
final_answer ([`~agents.FinalAnswerPromptTemplate`]): Final answer prompt templates.
"""
system_prompt: str
planning: PlanningPromptTemplate
managed_agent: ManagedAgentPromptTemplate
final_answer: FinalAnswerPromptTemplate
EMPTY_PROMPT_TEMPLATES = PromptTemplates(
system_prompt="",
planning=PlanningPromptTemplate(
initial_plan="",
update_plan_pre_messages="",
update_plan_post_messages="",
),
managed_agent=ManagedAgentPromptTemplate(task="", report=""),
final_answer=FinalAnswerPromptTemplate(pre_messages="", post_messages=""),
)
@dataclass
class RunResult:
"""Holds extended information about an agent run.
Attributes:
output (Any | None): The final output of the agent run, if available.
state (Literal["success", "max_steps_error"]): The final state of the agent after the run.
messages (list[dict]): The agent's memory, as a list of messages.
token_usage (TokenUsage | None): Count of tokens used during the run.
timing (Timing): Timing details of the agent run: start time, end time, duration.
"""
output: Any | None
state: Literal["success", "max_steps_error"]
messages: list[dict]
token_usage: TokenUsage | None
timing: Timing
StreamEvent: TypeAlias = Union[
ChatMessageStreamDelta,
ChatMessageToolCall,
ActionOutput,
ToolCall,
ToolOutput,
PlanningStep,
ActionStep,
FinalAnswerStep,
]
class MultiStepAgent(ABC):
"""
Agent class that solves the given task step by step, using the ReAct framework:
While the objective is not reached, the agent will perform a cycle of action (given by the LLM) and observation (obtained from the environment).
Args:
tools (`list[Tool]`): [`Tool`]s that the agent can use.
model (`Callable[[list[dict[str, str]]], ChatMessage]`): Model that will generate the agent's actions.
prompt_templates ([`~agents.PromptTemplates`], *optional*): Prompt templates.
instructions (`str`, *optional*): Custom instructions for the agent, will be inserted in the system prompt.
max_steps (`int`, default `20`): Maximum number of steps the agent can take to solve the task.
add_base_tools (`bool`, default `False`): Whether to add the base tools to the agent's tools.
verbosity_level (`LogLevel`, default `LogLevel.INFO`): Level of verbosity of the agent's logs.
grammar (`dict[str, str]`, *optional*): Grammar used to parse the LLM output.
Parameter `grammar` is deprecated and will be removed in version 1.20.
managed_agents (`list`, *optional*): Managed agents that the agent can call.
step_callbacks (`list[Callable]`, *optional*): Callbacks that will be called at each step.
planning_interval (`int`, *optional*): Interval at which the agent will run a planning step.
name (`str`, *optional*): Necessary for a managed agent only - the name by which this agent can be called.
description (`str`, *optional*): Necessary for a managed agent only - the description of this agent.
provide_run_summary (`bool`, *optional*): Whether to provide a run summary when called as a managed agent.
final_answer_checks (`list[Callable]`, *optional*): List of validation functions to run before accepting a final answer.
Each function should:
- Take the final answer and the agent's memory as arguments.
- Return a boolean indicating whether the final answer is valid.
"""
def __init__(
self,
tools: list[Tool],
model: Model,
prompt_templates: PromptTemplates | None = None,
instructions: str | None = None,
max_steps: int = 20,
add_base_tools: bool = False,
verbosity_level: LogLevel = LogLevel.INFO,
grammar: dict[str, str] | None = None,
managed_agents: list | None = None,
step_callbacks: list[Callable] | None = None,
planning_interval: int | None = None,
name: str | None = None,
description: str | None = None,
provide_run_summary: bool = False,
final_answer_checks: list[Callable] | None = None,
return_full_result: bool = False,
logger: AgentLogger | None = None,
):
self.agent_name = self.__class__.__name__
self.model = model
self.prompt_templates = prompt_templates or EMPTY_PROMPT_TEMPLATES
if prompt_templates is not None:
missing_keys = set(EMPTY_PROMPT_TEMPLATES.keys()) - set(prompt_templates.keys())
assert not missing_keys, (
f"Some prompt templates are missing from your custom `prompt_templates`: {missing_keys}"
)
for key, value in EMPTY_PROMPT_TEMPLATES.items():
if isinstance(value, dict):
for subkey in value.keys():
assert key in prompt_templates.keys() and (subkey in prompt_templates[key].keys()), (
f"Some prompt templates are missing from your custom `prompt_templates`: {subkey} under {key}"
)
self.max_steps = max_steps
self.step_number = 0
if grammar is not None:
warnings.warn(
"Parameter 'grammar' is deprecated and will be removed in version 1.20.",
FutureWarning,
)
self.grammar = grammar
self.planning_interval = planning_interval
self.state: dict[str, Any] = {}
self.name = self._validate_name(name)
self.description = description
self.provide_run_summary = provide_run_summary
self.final_answer_checks = final_answer_checks if final_answer_checks is not None else []
self.return_full_result = return_full_result
self.instructions = instructions
self._setup_managed_agents(managed_agents)
self._setup_tools(tools, add_base_tools)
self._validate_tools_and_managed_agents(tools, managed_agents)
self.task: str | None = None
self.memory = AgentMemory(self.system_prompt)
if logger is None:
self.logger = AgentLogger(level=verbosity_level)
else:
self.logger = logger
self.monitor = Monitor(self.model, self.logger)
self.step_callbacks = step_callbacks if step_callbacks is not None else []
self.step_callbacks.append(self.monitor.update_metrics)
self.stream_outputs = False
@property
def system_prompt(self) -> str:
return self.initialize_system_prompt()
@system_prompt.setter
def system_prompt(self, value: str):
raise AttributeError(
"""The 'system_prompt' property is read-only. Use 'self.prompt_templates["system_prompt"]' instead."""
)
def _validate_name(self, name: str | None) -> str | None:
if name is not None and not is_valid_name(name):
raise ValueError(f"Agent name '{name}' must be a valid Python identifier and not a reserved keyword.")
return name
def _setup_managed_agents(self, managed_agents: list | None = None) -> None:
"""Setup managed agents with proper logging."""
self.managed_agents = {}
if managed_agents:
assert all(agent.name and agent.description for agent in managed_agents), (
"All managed agents need both a name and a description!"
)
self.managed_agents = {agent.name: agent for agent in managed_agents}
# Ensure managed agents can be called as tools by the model: set their inputs and output_type
for agent in self.managed_agents.values():
agent.inputs = {
"task": {"type": "string", "description": "Long detailed description of the task."},
"additional_args": {
"type": "object",
"description": "Dictionary of extra inputs to pass to the managed agent, e.g. images, dataframes, or any other contextual data it may need.",
},
}
agent.output_type = "string"
def _setup_tools(self, tools, add_base_tools):
assert all(isinstance(tool, Tool) for tool in tools), "All elements must be instance of Tool (or a subclass)"
self.tools = {tool.name: tool for tool in tools}
if add_base_tools:
self.tools.update(
{
name: cls()
for name, cls in TOOL_MAPPING.items()
if name != "python_interpreter" or self.__class__.__name__ == "ToolCallingAgent"
}
)
self.tools.setdefault("final_answer", FinalAnswerTool())
def _validate_tools_and_managed_agents(self, tools, managed_agents):
tool_and_managed_agent_names = [tool.name for tool in tools]
if managed_agents is not None:
tool_and_managed_agent_names += [agent.name for agent in managed_agents]
if self.name:
tool_and_managed_agent_names.append(self.name)
if len(tool_and_managed_agent_names) != len(set(tool_and_managed_agent_names)):
raise ValueError(
"Each tool or managed_agent should have a unique name! You passed these duplicate names: "
f"{[name for name in tool_and_managed_agent_names if tool_and_managed_agent_names.count(name) > 1]}"
)
def run(
self,
task: str,
stream: bool = False,
reset: bool = True,
images: list["PIL.Image.Image"] | None = None,
additional_args: dict | None = None,
max_steps: int | None = None,
):
"""
Run the agent for the given task.
Args:
task (`str`): Task to perform.
stream (`bool`): Whether to run in streaming mode.
If `True`, returns a generator that yields each step as it is executed. You must iterate over this generator to process the individual steps (e.g., using a for loop or `next()`).
If `False`, executes all steps internally and returns only the final answer after completion.
reset (`bool`): Whether to reset the conversation or keep it going from previous run.
images (`list[PIL.Image.Image]`, *optional*): Image(s) objects.
additional_args (`dict`, *optional*): Any other variables that you want to pass to the agent run, for instance images or dataframes. Give them clear names!
max_steps (`int`, *optional*): Maximum number of steps the agent can take to solve the task. if not provided, will use the agent's default value.
Example:
```py
from smolagents import CodeAgent
agent = CodeAgent(tools=[])
agent.run("What is the result of 2 power 3.7384?")
```
"""
max_steps = max_steps or self.max_steps
self.task = task
self.interrupt_switch = False
if additional_args is not None:
self.state.update(additional_args)
self.task += f"""
You have been provided with these additional arguments, that you can access using the keys as variables in your python code:
{str(additional_args)}."""
self.memory.system_prompt = SystemPromptStep(system_prompt=self.system_prompt)
if reset:
self.memory.reset()
self.monitor.reset()
self.logger.log_task(
content=self.task.strip(),
subtitle=f"{type(self.model).__name__} - {(self.model.model_id if hasattr(self.model, 'model_id') else '')}",
level=LogLevel.INFO,
title=self.name if hasattr(self, "name") else None,
)
self.memory.steps.append(TaskStep(task=self.task, task_images=images))
if getattr(self, "python_executor", None):
self.python_executor.send_variables(variables=self.state)
self.python_executor.send_tools({**self.tools, **self.managed_agents})
if stream:
# The steps are returned as they are executed through a generator to iterate on.
return self._run_stream(task=self.task, max_steps=max_steps, images=images)
run_start_time = time.time()
# Outputs are returned only at the end. We only look at the last step.
steps = list(self._run_stream(task=self.task, max_steps=max_steps, images=images))
assert isinstance(steps[-1], FinalAnswerStep)
output = steps[-1].output
if self.return_full_result:
total_input_tokens = 0
total_output_tokens = 0
correct_token_usage = True
for step in self.memory.steps:
if isinstance(step, (ActionStep, PlanningStep)):
if step.token_usage is None:
correct_token_usage = False
break
else:
total_input_tokens += step.token_usage.input_tokens
total_output_tokens += step.token_usage.output_tokens
if correct_token_usage:
token_usage = TokenUsage(input_tokens=total_input_tokens, output_tokens=total_output_tokens)
else:
token_usage = None
if self.memory.steps and isinstance(getattr(self.memory.steps[-1], "error", None), AgentMaxStepsError):
state = "max_steps_error"
else:
state = "success"
messages = self.memory.get_full_steps()
return RunResult(
output=output,
token_usage=token_usage,
messages=messages,
timing=Timing(start_time=run_start_time, end_time=time.time()),
state=state,
)
return output
def _run_stream(
self, task: str, max_steps: int, images: list["PIL.Image.Image"] | None = None
) -> Generator[ActionStep | PlanningStep | FinalAnswerStep | ChatMessageStreamDelta]:
self.step_number = 1
returned_final_answer = False
while not returned_final_answer and self.step_number <= max_steps:
if self.interrupt_switch:
raise AgentError("Agent interrupted.", self.logger)
# Run a planning step if scheduled
if self.planning_interval is not None and (
self.step_number == 1 or (self.step_number - 1) % self.planning_interval == 0
):
planning_start_time = time.time()
planning_step = None
for element in self._generate_planning_step(
task, is_first_step=len(self.memory.steps) == 1, step=self.step_number
): # Don't use the attribute step_number here, because there can be steps from previous runs
yield element
planning_step = element
assert isinstance(planning_step, PlanningStep) # Last yielded element should be a PlanningStep
self.memory.steps.append(planning_step)
planning_end_time = time.time()
planning_step.timing = Timing(
start_time=planning_start_time,
end_time=planning_end_time,
)
# Start action step!
action_step_start_time = time.time()
action_step = ActionStep(
step_number=self.step_number,
timing=Timing(start_time=action_step_start_time),
observations_images=images,
)
self.logger.log_rule(f"Step {self.step_number}", level=LogLevel.INFO)
try:
for output in self._step_stream(action_step):
# Yield all
yield output
if isinstance(output, ActionOutput) and output.is_final_answer:
final_answer = output.output
self.logger.log(
Text(f"Final answer: {final_answer}", style=f"bold {YELLOW_HEX}"),
level=LogLevel.INFO,
)
if self.final_answer_checks:
self._validate_final_answer(final_answer)
returned_final_answer = True
action_step.is_final_answer = True
except AgentGenerationError as e:
# Agent generation errors are not caused by a Model error but an implementation error: so we should raise them and exit.
raise e
except AgentError as e:
# Other AgentError types are caused by the Model, so we should log them and iterate.
action_step.error = e
finally:
self._finalize_step(action_step)
self.memory.steps.append(action_step)
yield action_step
self.step_number += 1
if not returned_final_answer and self.step_number == max_steps + 1:
final_answer = self._handle_max_steps_reached(task, images)
yield action_step
yield FinalAnswerStep(handle_agent_output_types(final_answer))
def _validate_final_answer(self, final_answer: Any):
for check_function in self.final_answer_checks:
try:
assert check_function(final_answer, self.memory)
except Exception as e:
raise AgentError(f"Check {check_function.__name__} failed with error: {e}", self.logger)
def _finalize_step(self, memory_step: ActionStep):
memory_step.timing.end_time = time.time()
for callback in self.step_callbacks:
# For compatibility with old callbacks that don't take the agent as an argument
callback(memory_step) if len(inspect.signature(callback).parameters) == 1 else callback(
memory_step, agent=self
)
def _handle_max_steps_reached(self, task: str, images: list["PIL.Image.Image"]) -> Any:
action_step_start_time = time.time()
final_answer = self.provide_final_answer(task, images)
final_memory_step = ActionStep(
step_number=self.step_number,
error=AgentMaxStepsError("Reached max steps.", self.logger),
timing=Timing(start_time=action_step_start_time, end_time=time.time()),
token_usage=final_answer.token_usage,
)
final_memory_step.action_output = final_answer.content
self._finalize_step(final_memory_step)
self.memory.steps.append(final_memory_step)
return final_answer.content
def _generate_planning_step(
self, task, is_first_step: bool, step: int
) -> Generator[ChatMessageStreamDelta | PlanningStep]:
start_time = time.time()
if is_first_step:
input_messages = [
ChatMessage(
role=MessageRole.USER,
content=[
{
"type": "text",
"text": populate_template(
self.prompt_templates["planning"]["initial_plan"],
variables={"task": task, "tools": self.tools, "managed_agents": self.managed_agents},
),
}
],
)
]
if self.stream_outputs and hasattr(self.model, "generate_stream"):
plan_message_content = ""
output_stream = self.model.generate_stream(input_messages, stop_sequences=[""]) # type: ignore
input_tokens, output_tokens = 0, 0
with Live("", console=self.logger.console, vertical_overflow="visible") as live:
for event in output_stream:
if event.content is not None:
plan_message_content += event.content
live.update(Markdown(plan_message_content))
if event.token_usage:
output_tokens += event.token_usage.output_tokens
input_tokens = event.token_usage.input_tokens
yield event
else:
plan_message = self.model.generate(input_messages, stop_sequences=[""])
plan_message_content = plan_message.content
input_tokens, output_tokens = (
(
plan_message.token_usage.input_tokens,
plan_message.token_usage.output_tokens,
)
if plan_message.token_usage
else (None, None)
)
plan = textwrap.dedent(
f"""Here are the facts I know and the plan of action that I will follow to solve the task:\n```\n{plan_message_content}\n```"""
)
else:
# Summary mode removes the system prompt and previous planning messages output by the model.
# Removing previous planning messages avoids influencing too much the new plan.
memory_messages = self.write_memory_to_messages(summary_mode=True)
plan_update_pre = ChatMessage(
role=MessageRole.SYSTEM,
content=[
{
"type": "text",
"text": populate_template(
self.prompt_templates["planning"]["update_plan_pre_messages"], variables={"task": task}
),
}
],
)
plan_update_post = ChatMessage(
role=MessageRole.USER,
content=[
{
"type": "text",
"text": populate_template(
self.prompt_templates["planning"]["update_plan_post_messages"],
variables={
"task": task,
"tools": self.tools,
"managed_agents": self.managed_agents,
"remaining_steps": (self.max_steps - step),
},
),
}
],
)
input_messages = [plan_update_pre] + memory_messages + [plan_update_post]
if self.stream_outputs and hasattr(self.model, "generate_stream"):
plan_message_content = ""
input_tokens, output_tokens = 0, 0
with Live("", console=self.logger.console, vertical_overflow="visible") as live:
for event in self.model.generate_stream(
input_messages,
stop_sequences=[""],
): # type: ignore
if event.content is not None:
plan_message_content += event.content
live.update(Markdown(plan_message_content))
if event.token_usage:
output_tokens += event.token_usage.output_tokens
input_tokens = event.token_usage.input_tokens
yield event
else:
plan_message = self.model.generate(input_messages, stop_sequences=[""])
plan_message_content = plan_message.content
if plan_message.token_usage is not None:
input_tokens, output_tokens = (
plan_message.token_usage.input_tokens,
plan_message.token_usage.output_tokens,
)
plan = textwrap.dedent(
f"""I still need to solve the task I was given:\n```\n{self.task}\n```\n\nHere are the facts I know and my new/updated plan of action to solve the task:\n```\n{plan_message_content}\n```"""
)
log_headline = "Initial plan" if is_first_step else "Updated plan"
self.logger.log(Rule(f"[bold]{log_headline}", style="orange"), Text(plan), level=LogLevel.INFO)
yield PlanningStep(
model_input_messages=input_messages,
plan=plan,
model_output_message=ChatMessage(role=MessageRole.ASSISTANT, content=plan_message_content),
token_usage=TokenUsage(input_tokens=input_tokens, output_tokens=output_tokens),
timing=Timing(start_time=start_time, end_time=time.time()),
)
@property
def logs(self):
logger.warning(
"The 'logs' attribute is deprecated and will soon be removed. Please use 'self.memory.steps' instead."
)
return [self.memory.system_prompt] + self.memory.steps
@abstractmethod
def initialize_system_prompt(self) -> str:
"""To be implemented in child classes"""
...
def interrupt(self):
"""Interrupts the agent execution."""
self.interrupt_switch = True
def write_memory_to_messages(
self,
summary_mode: bool = False,
) -> list[ChatMessage]:
"""
Reads past llm_outputs, actions, and observations or errors from the memory into a series of messages
that can be used as input to the LLM. Adds a number of keywords (such as PLAN, error, etc) to help
the LLM.
"""
messages = self.memory.system_prompt.to_messages(summary_mode=summary_mode)
for memory_step in self.memory.steps:
messages.extend(memory_step.to_messages(summary_mode=summary_mode))
return messages
def _step_stream(
self, memory_step: ActionStep
) -> Generator[ChatMessageStreamDelta | ToolCall | ToolOutput | ActionOutput]:
"""
Perform one step in the ReAct framework: the agent thinks, acts, and observes the result.
Yields ChatMessageStreamDelta during the run if streaming is enabled.
At the end, yields either None if the step is not final, or the final answer.
"""
raise NotImplementedError("This method should be implemented in child classes")
def step(self, memory_step: ActionStep) -> Any:
"""
Perform one step in the ReAct framework: the agent thinks, acts, and observes the result.
Returns either None if the step is not final, or the final answer.
"""
return list(self._step_stream(memory_step))[-1]
def extract_action(self, model_output: str, split_token: str) -> tuple[str, str]:
"""
Parse action from the LLM output
Args:
model_output (`str`): Output of the LLM
split_token (`str`): Separator for the action. Should match the example in the system prompt.
"""
try:
split = model_output.split(split_token)
rationale, action = (
split[-2],
split[-1],
) # NOTE: using indexes starting from the end solves for when you have more than one split_token in the output
except Exception:
raise AgentParsingError(
f"No '{split_token}' token provided in your output.\nYour output:\n{model_output}\n. Be sure to include an action, prefaced with '{split_token}'!",
self.logger,
)
return rationale.strip(), action.strip()
def provide_final_answer(self, task: str, images: list["PIL.Image.Image"] | None = None) -> ChatMessage:
"""
Provide the final answer to the task, based on the logs of the agent's interactions.
Args:
task (`str`): Task to perform.
images (`list[PIL.Image.Image]`, *optional*): Image(s) objects.
Returns:
`str`: Final answer to the task.
"""
messages = [
ChatMessage(
role=MessageRole.SYSTEM,
content=[
{
"type": "text",
"text": self.prompt_templates["final_answer"]["pre_messages"],
}
],
)
]
if images:
messages[0].content += [{"type": "image", "image": image} for image in images]
messages += self.write_memory_to_messages()[1:]
messages.append(
ChatMessage(
role=MessageRole.USER,
content=[
{
"type": "text",
"text": populate_template(
self.prompt_templates["final_answer"]["post_messages"], variables={"task": task}
),
}
],
)
)
try:
chat_message: ChatMessage = self.model.generate(messages)
return chat_message
except Exception as e:
return ChatMessage(role=MessageRole.ASSISTANT, content=f"Error in generating final LLM output:\n{e}")
def visualize(self):
"""Creates a rich tree visualization of the agent's structure."""
self.logger.visualize_agent_tree(self)
def replay(self, detailed: bool = False):
"""Prints a pretty replay of the agent's steps.
Args:
detailed (bool, optional): If True, also displays the memory at each step. Defaults to False.
Careful: will increase log length exponentially. Use only for debugging.
"""
self.memory.replay(self.logger, detailed=detailed)
def __call__(self, task: str, **kwargs):
"""Adds additional prompting for the managed agent, runs it, and wraps the output.
This method is called only by a managed agent.
"""
full_task = populate_template(
self.prompt_templates["managed_agent"]["task"],
variables=dict(name=self.name, task=task),
)
result = self.run(full_task, **kwargs)
if isinstance(result, RunResult):
report = result.output
else:
report = result
answer = populate_template(
self.prompt_templates["managed_agent"]["report"], variables=dict(name=self.name, final_answer=report)
)
if self.provide_run_summary:
answer += "\n\nFor more detail, find below a summary of this agent's work:\n\n"
for message in self.write_memory_to_messages(summary_mode=True):
content = message["content"]
answer += "\n" + truncate_content(str(content)) + "\n---"
answer += "\n"
return answer
def save(self, output_dir: str | Path, relative_path: str | None = None):
"""
Saves the relevant code files for your agent. This will copy the code of your agent in `output_dir` as well as autogenerate:
- a `tools` folder containing the logic for each of the tools under `tools/{tool_name}.py`.
- a `managed_agents` folder containing the logic for each of the managed agents.
- an `agent.json` file containing a dictionary representing your agent.
- a `prompt.yaml` file containing the prompt templates used by your agent.
- an `app.py` file providing a UI for your agent when it is exported to a Space with `agent.push_to_hub()`
- a `requirements.txt` containing the names of the modules used by your tool (as detected when inspecting its
code)
Args:
output_dir (`str` or `Path`): The folder in which you want to save your agent.
"""
make_init_file(output_dir)
# Recursively save managed agents
if self.managed_agents:
make_init_file(os.path.join(output_dir, "managed_agents"))
for agent_name, agent in self.managed_agents.items():
agent_suffix = f"managed_agents.{agent_name}"
if relative_path:
agent_suffix = relative_path + "." + agent_suffix
agent.save(os.path.join(output_dir, "managed_agents", agent_name), relative_path=agent_suffix)
class_name = self.__class__.__name__
# Save tools to different .py files
for tool in self.tools.values():
make_init_file(os.path.join(output_dir, "tools"))
tool.save(os.path.join(output_dir, "tools"), tool_file_name=tool.name, make_gradio_app=False)
# Save prompts to yaml
yaml_prompts = yaml.safe_dump(
self.prompt_templates,
default_style="|", # This forces block literals for all strings
default_flow_style=False,
width=float("inf"),
sort_keys=False,
allow_unicode=True,
indent=2,
)
with open(os.path.join(output_dir, "prompts.yaml"), "w", encoding="utf-8") as f:
f.write(yaml_prompts)
# Save agent dictionary to json
agent_dict = self.to_dict()
agent_dict["tools"] = [tool.name for tool in self.tools.values()]
agent_dict["managed_agents"] = {agent.name: agent.__class__.__name__ for agent in self.managed_agents.values()}
with open(os.path.join(output_dir, "agent.json"), "w", encoding="utf-8") as f:
json.dump(agent_dict, f, indent=4)
# Save requirements
with open(os.path.join(output_dir, "requirements.txt"), "w", encoding="utf-8") as f:
f.writelines(f"{r}\n" for r in agent_dict["requirements"])
# Make agent.py file with Gradio UI
agent_name = f"agent_{self.name}" if getattr(self, "name", None) else "agent"
managed_agent_relative_path = relative_path + "." if relative_path is not None else ""
app_template = AGENT_GRADIO_APP_TEMPLATE
template_env = jinja2.Environment(loader=jinja2.BaseLoader(), undefined=jinja2.StrictUndefined)
template_env.filters["repr"] = repr
template_env.filters["camelcase"] = lambda value: "".join(word.capitalize() for word in value.split("_"))
template = template_env.from_string(app_template)
# Render the app.py file from Jinja2 template
app_text = template.render(
{
"agent_name": agent_name,
"class_name": class_name,
"agent_dict": agent_dict,
"tools": self.tools,
"managed_agents": self.managed_agents,
"managed_agent_relative_path": managed_agent_relative_path,
}
)
with open(os.path.join(output_dir, "app.py"), "w", encoding="utf-8") as f:
f.write(app_text + "\n") # Append newline at the end
def to_dict(self) -> dict[str, Any]:
"""Convert the agent to a dictionary representation.
Returns:
`dict`: Dictionary representation of the agent.
"""
# TODO: handle serializing step_callbacks and final_answer_checks
for attr in ["final_answer_checks", "step_callbacks"]:
if getattr(self, attr, None):
self.logger.log(f"This agent has {attr}: they will be ignored by this method.", LogLevel.INFO)
tool_dicts = [tool.to_dict() for tool in self.tools.values()]
tool_requirements = {req for tool in self.tools.values() for req in tool.to_dict()["requirements"]}
managed_agents_requirements = {
req for managed_agent in self.managed_agents.values() for req in managed_agent.to_dict()["requirements"]
}
requirements = tool_requirements | managed_agents_requirements
if hasattr(self, "authorized_imports"):
requirements.update(
{package.split(".")[0] for package in self.authorized_imports if package not in BASE_BUILTIN_MODULES}
)
agent_dict = {
"class": self.__class__.__name__,
"tools": tool_dicts,
"model": {
"class": self.model.__class__.__name__,
"data": self.model.to_dict(),
},
"managed_agents": [managed_agent.to_dict() for managed_agent in self.managed_agents.values()],
"prompt_templates": self.prompt_templates,
"max_steps": self.max_steps,
"verbosity_level": int(self.logger.level),
"grammar": self.grammar,
"planning_interval": self.planning_interval,
"name": self.name,
"description": self.description,
"requirements": sorted(requirements),
}
return agent_dict
@classmethod
def from_dict(cls, agent_dict: dict[str, Any], **kwargs) -> "MultiStepAgent":
"""Create agent from a dictionary representation.
Args:
agent_dict (`dict[str, Any]`): Dictionary representation of the agent.
**kwargs: Additional keyword arguments that will override agent_dict values.
Returns:
`MultiStepAgent`: Instance of the agent class.
"""
# Load model
model_info = agent_dict["model"]
model_class = getattr(importlib.import_module("smolagents.models"), model_info["class"])
model = model_class.from_dict(model_info["data"])
# Load tools
tools = []
for tool_info in agent_dict["tools"]:
tools.append(Tool.from_code(tool_info["code"]))
# Load managed agents
managed_agents = []
for managed_agent_name, managed_agent_class_name in agent_dict["managed_agents"].items():
managed_agent_class = getattr(importlib.import_module("smolagents.agents"), managed_agent_class_name)
managed_agents.append(managed_agent_class.from_dict(agent_dict["managed_agents"][managed_agent_name]))
# Extract base agent parameters
agent_args = {
"model": model,
"tools": tools,
"prompt_templates": agent_dict.get("prompt_templates"),
"max_steps": agent_dict.get("max_steps"),
"verbosity_level": agent_dict.get("verbosity_level"),
"grammar": agent_dict.get("grammar"),
"planning_interval": agent_dict.get("planning_interval"),
"name": agent_dict.get("name"),
"description": agent_dict.get("description"),
}
# Filter out None values to use defaults from __init__
agent_args = {k: v for k, v in agent_args.items() if v is not None}
# Update with any additional kwargs
agent_args.update(kwargs)
# Create agent instance
return cls(**agent_args)
@classmethod
def from_hub(
cls,
repo_id: str,
token: str | None = None,
trust_remote_code: bool = False,
**kwargs,
):
"""
Loads an agent defined on the Hub.
Loading a tool from the Hub means that you'll download the tool and execute it locally.
ALWAYS inspect the tool you're downloading before loading it within your runtime, as you would do when
installing a package using pip/npm/apt.
Args:
repo_id (`str`):
The name of the repo on the Hub where your tool is defined.
token (`str`, *optional*):
The token to identify you on hf.co. If unset, will use the token generated when running
`huggingface-cli login` (stored in `~/.huggingface`).
trust_remote_code(`bool`, *optional*, defaults to False):
This flags marks that you understand the risk of running remote code and that you trust this tool.
If not setting this to True, loading the tool from Hub will fail.
kwargs (additional keyword arguments, *optional*):
Additional keyword arguments that will be split in two: all arguments relevant to the Hub (such as
`cache_dir`, `revision`, `subfolder`) will be used when downloading the files for your agent, and the
others will be passed along to its init.
"""
if not trust_remote_code:
raise ValueError(
"Loading an agent from Hub requires to acknowledge you trust its code: to do so, pass `trust_remote_code=True`."
)
# Get the agent's Hub folder.
download_kwargs = {"token": token, "repo_type": "space"} | {
key: kwargs.pop(key)
for key in [
"cache_dir",
"force_download",
"proxies",
"revision",
"local_files_only",
]
if key in kwargs
}
download_folder = Path(snapshot_download(repo_id=repo_id, **download_kwargs))
return cls.from_folder(download_folder, **kwargs)
@classmethod
def from_folder(cls, folder: str | Path, **kwargs):
"""Loads an agent from a local folder.
Args:
folder (`str` or `Path`): The folder where the agent is saved.
**kwargs: Additional keyword arguments that will be passed to the agent's init.
"""
# Load agent.json
folder = Path(folder)
agent_dict = json.loads((folder / "agent.json").read_text())
# Load managed agents from their respective folders, recursively
managed_agents = []
for managed_agent_name, managed_agent_class_name in agent_dict["managed_agents"].items():
agent_cls = getattr(importlib.import_module("smolagents.agents"), managed_agent_class_name)
managed_agents.append(agent_cls.from_folder(folder / "managed_agents" / managed_agent_name))
agent_dict["managed_agents"] = {}
# Load tools
tools = []
for tool_name in agent_dict["tools"]:
tool_code = (folder / "tools" / f"{tool_name}.py").read_text()
tools.append({"name": tool_name, "code": tool_code})
agent_dict["tools"] = tools
# Add managed agents to kwargs to override the empty list in from_dict
if managed_agents:
kwargs["managed_agents"] = managed_agents
return cls.from_dict(agent_dict, **kwargs)
def push_to_hub(
self,
repo_id: str,
commit_message: str = "Upload agent",
private: bool | None = None,
token: bool | str | None = None,
create_pr: bool = False,
) -> str:
"""
Upload the agent to the Hub.
Parameters:
repo_id (`str`):
The name of the repository you want to push to. It should contain your organization name when
pushing to a given organization.
commit_message (`str`, *optional*, defaults to `"Upload agent"`):
Message to commit while pushing.
private (`bool`, *optional*, defaults to `None`):
Whether to make the repo private. If `None`, the repo will be public unless the organization's default is private. This value is ignored if the repo already exists.
token (`bool` or `str`, *optional*):
The token to use as HTTP bearer authorization for remote files. If unset, will use the token generated
when running `huggingface-cli login` (stored in `~/.huggingface`).
create_pr (`bool`, *optional*, defaults to `False`):
Whether to create a PR with the uploaded files or directly commit.
"""
repo_url = create_repo(
repo_id=repo_id,
token=token,
private=private,
exist_ok=True,
repo_type="space",
space_sdk="gradio",
)
repo_id = repo_url.repo_id
metadata_update(
repo_id,
{"tags": ["smolagents", "agent"]},
repo_type="space",
token=token,
overwrite=True,
)
with tempfile.TemporaryDirectory() as work_dir:
self.save(work_dir)
logger.info(f"Uploading the following files to {repo_id}: {','.join(os.listdir(work_dir))}")
return upload_folder(
repo_id=repo_id,
commit_message=commit_message,
folder_path=work_dir,
token=token,
create_pr=create_pr,
repo_type="space",
)
class ToolCallingAgent(MultiStepAgent):
"""
This agent uses JSON-like tool calls, using method `model.get_tool_call` to leverage the LLM engine's tool calling capabilities.
Args:
tools (`list[Tool]`): [`Tool`]s that the agent can use.
model (`Model`): Model that will generate the agent's actions.
prompt_templates ([`~agents.PromptTemplates`], *optional*): Prompt templates.
planning_interval (`int`, *optional*): Interval at which the agent will run a planning step.
stream_outputs (`bool`, *optional*, default `False`): Whether to stream outputs during execution.
max_tool_threads (`int`, *optional*): Maximum number of threads for parallel tool calls.
Higher values increase concurrency but resource usage as well.
Defaults to `ThreadPoolExecutor`'s default.
**kwargs: Additional keyword arguments.
"""
def __init__(
self,
tools: list[Tool],
model: Model,
prompt_templates: PromptTemplates | None = None,
planning_interval: int | None = None,
stream_outputs: bool = False,
max_tool_threads: int | None = None,
**kwargs,
):
prompt_templates = prompt_templates or yaml.safe_load(
importlib.resources.files("smolagents.prompts").joinpath("toolcalling_agent.yaml").read_text()
)
super().__init__(
tools=tools,
model=model,
prompt_templates=prompt_templates,
planning_interval=planning_interval,
**kwargs,
)
# Streaming setup
self.stream_outputs = stream_outputs
if self.stream_outputs and not hasattr(self.model, "generate_stream"):
raise ValueError(
"`stream_outputs` is set to True, but the model class implements no `generate_stream` method."
)
# Tool calling setup
self.max_tool_threads = max_tool_threads
@property
def tools_and_managed_agents(self):
"""Returns a combined list of tools and managed agents."""
return list(self.tools.values()) + list(self.managed_agents.values())
def initialize_system_prompt(self) -> str:
system_prompt = populate_template(
self.prompt_templates["system_prompt"],
variables={
"tools": self.tools,
"managed_agents": self.managed_agents,
"custom_instructions": self.instructions,
},
)
return system_prompt
def _step_stream(
self, memory_step: ActionStep
) -> Generator[ChatMessageStreamDelta | ToolCall | ToolOutput | ActionOutput]:
"""
Perform one step in the ReAct framework: the agent thinks, acts, and observes the result.
Yields ChatMessageStreamDelta during the run if streaming is enabled.
At the end, yields either None if the step is not final, or the final answer.
"""
memory_messages = self.write_memory_to_messages()
input_messages = memory_messages.copy()
# Add new step in logs
memory_step.model_input_messages = input_messages
try:
if self.stream_outputs and hasattr(self.model, "generate_stream"):
output_stream = self.model.generate_stream(
input_messages,
stop_sequences=["Observation:", "Calling tools:"],
tools_to_call_from=self.tools_and_managed_agents,
)
chat_message_stream_deltas: list[ChatMessageStreamDelta] = []
with Live("", console=self.logger.console, vertical_overflow="visible") as live:
for event in output_stream:
chat_message_stream_deltas.append(event)
live.update(
Markdown(agglomerate_stream_deltas(chat_message_stream_deltas).render_as_markdown())
)
yield event
chat_message = agglomerate_stream_deltas(chat_message_stream_deltas)
else:
chat_message: ChatMessage = self.model.generate(
input_messages,
stop_sequences=["Observation:", "Calling tools:"],
tools_to_call_from=self.tools_and_managed_agents,
)
if chat_message.content is None and chat_message.raw is not None:
log_content = str(chat_message.raw)
else:
log_content = str(chat_message.content) or ""
self.logger.log_markdown(
content=log_content,
title="Output message of the LLM:",
level=LogLevel.DEBUG,
)
# Record model output
memory_step.model_output_message = chat_message
memory_step.model_output = chat_message.content
memory_step.token_usage = chat_message.token_usage
except Exception as e:
raise AgentGenerationError(f"Error while generating output:\n{e}", self.logger) from e
if chat_message.tool_calls is None or len(chat_message.tool_calls) == 0:
try:
chat_message = self.model.parse_tool_calls(chat_message)
except Exception as e:
raise AgentParsingError(f"Error while parsing tool call from model output: {e}", self.logger)
else:
for tool_call in chat_message.tool_calls:
tool_call.function.arguments = parse_json_if_needed(tool_call.function.arguments)
final_answer, got_final_answer = None, False
for output in self.process_tool_calls(chat_message, memory_step):
yield output
if isinstance(output, ToolOutput):
if output.is_final_answer:
if got_final_answer:
raise AgentToolExecutionError(
"You returned multiple final answers. Please return only one single final answer!",
self.logger,
)
final_answer = output.output
got_final_answer = True
# Manage state variables
if isinstance(final_answer, str) and final_answer in self.state.keys():
final_answer = self.state[final_answer]
yield ActionOutput(
output=final_answer,
is_final_answer=got_final_answer,
)
def process_tool_calls(
self, chat_message: ChatMessage, memory_step: ActionStep
) -> Generator[ToolCall | ToolOutput]:
"""Process tool calls from the model output and update agent memory.
Args:
chat_message (`ChatMessage`): Chat message containing tool calls from the model.
memory_step (`ActionStep)`: Memory ActionStep to update with results.
Yields:
`ToolCall | ToolOutput`: The tool call or tool output.
"""
parallel_calls: dict[str, ToolCall] = {}
assert chat_message.tool_calls is not None
for chat_tool_call in chat_message.tool_calls:
tool_call = ToolCall(
name=chat_tool_call.function.name, arguments=chat_tool_call.function.arguments, id=chat_tool_call.id
)
yield tool_call
parallel_calls[tool_call.id] = tool_call
# Helper function to process a single tool call
def process_single_tool_call(tool_call: ToolCall) -> ToolOutput:
tool_name = tool_call.name
tool_arguments = tool_call.arguments or {}
self.logger.log(
Panel(Text(f"Calling tool: '{tool_name}' with arguments: {tool_arguments}")),
level=LogLevel.INFO,
)
tool_call_result = self.execute_tool_call(tool_name, tool_arguments)
tool_call_result_type = type(tool_call_result)
if tool_call_result_type in [AgentImage, AgentAudio]:
if tool_call_result_type == AgentImage:
observation_name = "image.png"
elif tool_call_result_type == AgentAudio:
observation_name = "audio.mp3"
# TODO: tool_call_result naming could allow for different names of same type
self.state[observation_name] = tool_call_result
observation = f"Stored '{observation_name}' in memory."
else:
observation = str(tool_call_result).strip()
self.logger.log(
f"Observations: {observation.replace('[', '|')}", # escape potential rich-tag-like components
level=LogLevel.INFO,
)
is_final_answer = tool_name == "final_answer"
return ToolOutput(
id=tool_call.id,
output=tool_call_result,
is_final_answer=is_final_answer,
observation=observation,
tool_call=tool_call,
)
# Process tool calls in parallel
outputs = {}
if len(parallel_calls) == 1:
# If there's only one call, process it directly
tool_call = list(parallel_calls.values())[0]
tool_output = process_single_tool_call(tool_call)
outputs[tool_output.id] = tool_output
yield tool_output
else:
# If multiple tool calls, process them in parallel
with ThreadPoolExecutor(self.max_tool_threads) as executor:
futures = [
executor.submit(process_single_tool_call, tool_call) for tool_call in parallel_calls.values()
]
for future in as_completed(futures):
tool_output = future.result()
outputs[tool_output.id] = tool_output
yield tool_output
memory_step.tool_calls = [parallel_calls[k] for k in sorted(parallel_calls.keys())]
memory_step.model_output = memory_step.model_output or ""
memory_step.observations = memory_step.observations or ""
for tool_output in [outputs[k] for k in sorted(outputs.keys())]:
message = f"Tool call {tool_output.id}: calling '{tool_output.tool_call.name}' with arguments: {tool_output.tool_call.arguments}\n"
memory_step.model_output += message
memory_step.observations += tool_output.observation + "\n"
memory_step.model_output = memory_step.model_output.rstrip("\n")
memory_step.observations = (
memory_step.observations.rstrip("\n") if memory_step.observations else memory_step.observations
)
def _substitute_state_variables(self, arguments: dict[str, str] | str) -> dict[str, Any] | str:
"""Replace string values in arguments with their corresponding state values if they exist."""
if isinstance(arguments, dict):
return {
key: self.state.get(value, value) if isinstance(value, str) else value
for key, value in arguments.items()
}
return arguments
def execute_tool_call(self, tool_name: str, arguments: dict[str, str] | str) -> Any:
"""
Execute a tool or managed agent with the provided arguments.
The arguments are replaced with the actual values from the state if they refer to state variables.
Args:
tool_name (`str`): Name of the tool or managed agent to execute.
arguments (dict[str, str] | str): Arguments passed to the tool call.
"""
# Check if the tool exists
available_tools = {**self.tools, **self.managed_agents}
if tool_name not in available_tools:
raise AgentToolExecutionError(
f"Unknown tool {tool_name}, should be one of: {', '.join(available_tools)}.", self.logger
)
# Get the tool and substitute state variables in arguments
tool = available_tools[tool_name]
arguments = self._substitute_state_variables(arguments)
is_managed_agent = tool_name in self.managed_agents
error_msg = validate_tool_arguments(tool, arguments)
if error_msg:
raise AgentToolCallError(error_msg, self.logger)
try:
# Call tool with appropriate arguments
if isinstance(arguments, dict):
return tool(**arguments) if is_managed_agent else tool(**arguments, sanitize_inputs_outputs=True)
else:
return tool(arguments) if is_managed_agent else tool(arguments, sanitize_inputs_outputs=True)
except Exception as e:
# Handle execution errors
if is_managed_agent:
error_msg = (
f"Error executing request to team member '{tool_name}' with arguments {str(arguments)}: {e}\n"
"Please try again or request to another team member"
)
else:
error_msg = (
f"Error executing tool '{tool_name}' with arguments {str(arguments)}: {type(e).__name__}: {e}\n"
"Please try again or use another tool"
)
raise AgentToolExecutionError(error_msg, self.logger) from e
class CodeAgent(MultiStepAgent):
"""
In this agent, the tool calls will be formulated by the LLM in code format, then parsed and executed.
Args:
tools (`list[Tool]`): [`Tool`]s that the agent can use.
model (`Model`): Model that will generate the agent's actions.
prompt_templates ([`~agents.PromptTemplates`], *optional*): Prompt templates.
additional_authorized_imports (`list[str]`, *optional*): Additional authorized imports for the agent.
planning_interval (`int`, *optional*): Interval at which the agent will run a planning step.
executor_type (`str`, default `"local"`): Which executor type to use between `"local"`, `"e2b"`, or `"docker"`.
executor_kwargs (`dict`, *optional*): Additional arguments to pass to initialize the executor.
max_print_outputs_length (`int`, *optional*): Maximum length of the print outputs.
stream_outputs (`bool`, *optional*, default `False`): Whether to stream outputs during execution.
use_structured_outputs_internally (`bool`, default `False`): Whether to use structured generation at each action step: improves performance for many models.
grammar (`dict[str, str]`, *optional*): Grammar used to parse the LLM output.
Parameter `grammar` is deprecated and will be removed in version 1.20.
**kwargs: Additional keyword arguments.
"""
def __init__(
self,
tools: list[Tool],
model: Model,
prompt_templates: PromptTemplates | None = None,
additional_authorized_imports: list[str] | None = None,
planning_interval: int | None = None,
executor_type: str | None = "local",
executor_kwargs: dict[str, Any] | None = None,
max_print_outputs_length: int | None = None,
stream_outputs: bool = False,
use_structured_outputs_internally: bool = False,
grammar: dict[str, str] | None = None,
**kwargs,
):
self.additional_authorized_imports = additional_authorized_imports if additional_authorized_imports else []
self.authorized_imports = sorted(set(BASE_BUILTIN_MODULES) | set(self.additional_authorized_imports))
self.max_print_outputs_length = max_print_outputs_length
self._use_structured_outputs_internally = use_structured_outputs_internally
if use_structured_outputs_internally:
prompt_templates = prompt_templates or yaml.safe_load(
importlib.resources.files("smolagents.prompts").joinpath("structured_code_agent.yaml").read_text()
)
else:
prompt_templates = prompt_templates or yaml.safe_load(
importlib.resources.files("smolagents.prompts").joinpath("code_agent.yaml").read_text()
)
if grammar and use_structured_outputs_internally:
raise ValueError("You cannot use 'grammar' and 'use_structured_outputs_internally' at the same time.")
super().__init__(
tools=tools,
model=model,
prompt_templates=prompt_templates,
grammar=grammar,
planning_interval=planning_interval,
**kwargs,
)
self.stream_outputs = stream_outputs
if self.stream_outputs and not hasattr(self.model, "generate_stream"):
raise ValueError(
"`stream_outputs` is set to True, but the model class implements no `generate_stream` method."
)
if "*" in self.additional_authorized_imports:
self.logger.log(
"Caution: you set an authorization for all imports, meaning your agent can decide to import any package it deems necessary. This might raise issues if the package is not installed in your environment.",
level=LogLevel.INFO,
)
self.executor_type = executor_type or "local"
self.executor_kwargs = executor_kwargs or {}
self.python_executor = self.create_python_executor()
def __enter__(self):
return self
def __exit__(self, exc_type, exc_value, traceback):
self.cleanup()
def cleanup(self):
"""Clean up resources used by the agent, such as the remote Python executor."""
if hasattr(self.python_executor, "cleanup"):
self.python_executor.cleanup()
def create_python_executor(self) -> PythonExecutor:
match self.executor_type:
case "e2b" | "docker":
if self.managed_agents:
raise Exception("Managed agents are not yet supported with remote code execution.")
if self.executor_type == "e2b":
return E2BExecutor(self.additional_authorized_imports, self.logger, **self.executor_kwargs)
else:
return DockerExecutor(self.additional_authorized_imports, self.logger, **self.executor_kwargs)
case "local":
return LocalPythonExecutor(
self.additional_authorized_imports,
**{"max_print_outputs_length": self.max_print_outputs_length} | self.executor_kwargs,
)
case _: # if applicable
raise ValueError(f"Unsupported executor type: {self.executor_type}")
def initialize_system_prompt(self) -> str:
system_prompt = populate_template(
self.prompt_templates["system_prompt"],
variables={
"tools": self.tools,
"managed_agents": self.managed_agents,
"authorized_imports": (
"You can import from any package you want."
if "*" in self.authorized_imports
else str(self.authorized_imports)
),
"custom_instructions": self.instructions,
},
)
return system_prompt
def _step_stream(
self, memory_step: ActionStep
) -> Generator[ChatMessageStreamDelta | ToolCall | ToolOutput | ActionOutput]:
"""
Perform one step in the ReAct framework: the agent thinks, acts, and observes the result.
Yields ChatMessageStreamDelta during the run if streaming is enabled.
At the end, yields either None if the step is not final, or the final answer.
"""
memory_messages = self.write_memory_to_messages()
input_messages = memory_messages.copy()
### Generate model output ###
memory_step.model_input_messages = input_messages
try:
additional_args: dict[str, Any] = {}
if self.grammar:
additional_args["grammar"] = self.grammar
if self._use_structured_outputs_internally:
additional_args["response_format"] = CODEAGENT_RESPONSE_FORMAT
if self.stream_outputs:
output_stream = self.model.generate_stream(
input_messages,
stop_sequences=["", "Observation:", "Calling tools:"],
**additional_args,
)
chat_message_stream_deltas: list[ChatMessageStreamDelta] = []
with Live("", console=self.logger.console, vertical_overflow="visible") as live:
for event in output_stream:
chat_message_stream_deltas.append(event)
live.update(
Markdown(agglomerate_stream_deltas(chat_message_stream_deltas).render_as_markdown())
)
yield event
chat_message = agglomerate_stream_deltas(chat_message_stream_deltas)
memory_step.model_output_message = chat_message
output_text = chat_message.content
else:
chat_message: ChatMessage = self.model.generate(
input_messages,
stop_sequences=["", "Observation:", "Calling tools:"],
**additional_args,
)
memory_step.model_output_message = chat_message
output_text = chat_message.content
self.logger.log_markdown(
content=output_text,
title="Output message of the LLM:",
level=LogLevel.DEBUG,
)
# This adds sequence to the history.
# This will nudge ulterior LLM calls to finish with , thus efficiently stopping generation.
if output_text and output_text.strip().endswith("```"):
output_text += ""
memory_step.model_output_message.content = output_text
memory_step.token_usage = chat_message.token_usage
memory_step.model_output = output_text
except Exception as e:
raise AgentGenerationError(f"Error in generating model output:\n{e}", self.logger) from e
### Parse output ###
try:
if self._use_structured_outputs_internally:
code_action = json.loads(output_text)["code"]
code_action = extract_code_from_text(code_action) or code_action
else:
code_action = parse_code_blobs(output_text)
code_action = fix_final_answer_code(code_action)
memory_step.code_action = code_action
except Exception as e:
error_msg = f"Error in code parsing:\n{e}\nMake sure to provide correct code blobs."
raise AgentParsingError(error_msg, self.logger)
tool_call = ToolCall(
name="python_interpreter",
arguments=code_action,
id=f"call_{len(self.memory.steps)}",
)
yield tool_call
memory_step.tool_calls = [tool_call]
### Execute action ###
self.logger.log_code(title="Executing parsed code:", content=code_action, level=LogLevel.INFO)
is_final_answer = False
try:
output, execution_logs, is_final_answer = self.python_executor(code_action)
execution_outputs_console = []
if len(execution_logs) > 0:
execution_outputs_console += [
Text("Execution logs:", style="bold"),
Text(execution_logs),
]
observation = "Execution logs:\n" + execution_logs
except Exception as e:
if hasattr(self.python_executor, "state") and "_print_outputs" in self.python_executor.state:
execution_logs = str(self.python_executor.state["_print_outputs"])
if len(execution_logs) > 0:
execution_outputs_console = [
Text("Execution logs:", style="bold"),
Text(execution_logs),
]
memory_step.observations = "Execution logs:\n" + execution_logs
self.logger.log(Group(*execution_outputs_console), level=LogLevel.INFO)
error_msg = str(e)
if "Import of " in error_msg and " is not allowed" in error_msg:
self.logger.log(
"[bold red]Warning to user: Code execution failed due to an unauthorized import - Consider passing said import under `additional_authorized_imports` when initializing your CodeAgent.",
level=LogLevel.INFO,
)
raise AgentExecutionError(error_msg, self.logger)
truncated_output = truncate_content(str(output))
observation += "Last output from code snippet:\n" + truncated_output
memory_step.observations = observation
if not is_final_answer:
execution_outputs_console += [
Text(
f"Out: {truncated_output}",
),
]
self.logger.log(Group(*execution_outputs_console), level=LogLevel.INFO)
memory_step.action_output = output
yield ActionOutput(output=output, is_final_answer=is_final_answer)
def to_dict(self) -> dict[str, Any]:
"""Convert the agent to a dictionary representation.
Returns:
`dict`: Dictionary representation of the agent.
"""
agent_dict = super().to_dict()
agent_dict["authorized_imports"] = self.authorized_imports
agent_dict["executor_type"] = self.executor_type
agent_dict["executor_kwargs"] = self.executor_kwargs
agent_dict["max_print_outputs_length"] = self.max_print_outputs_length
return agent_dict
@classmethod
def from_dict(cls, agent_dict: dict[str, Any], **kwargs) -> "CodeAgent":
"""Create CodeAgent from a dictionary representation.
Args:
agent_dict (`dict[str, Any]`): Dictionary representation of the agent.
**kwargs: Additional keyword arguments that will override agent_dict values.
Returns:
`CodeAgent`: Instance of the CodeAgent class.
"""
# Add CodeAgent-specific parameters to kwargs
code_agent_kwargs = {
"additional_authorized_imports": agent_dict.get("authorized_imports"),
"executor_type": agent_dict.get("executor_type"),
"executor_kwargs": agent_dict.get("executor_kwargs"),
"max_print_outputs_length": agent_dict.get("max_print_outputs_length"),
}
# Filter out None values
code_agent_kwargs = {k: v for k, v in code_agent_kwargs.items() if v is not None}
# Update with any additional kwargs
code_agent_kwargs.update(kwargs)
# Call the parent class's from_dict method
return super().from_dict(agent_dict, **code_agent_kwargs)