agentscope.agents.react_agent module

An agent class that implements the ReAct algorithm. The agent will reason and act iteratively to solve problems. More details can be found in the paper https://arxiv.org/abs/2210.03629.

class agentscope.agents.react_agent.ReActAgent(*args: tuple, **kwargs: dict)[source]

Bases: AgentBase

An agent class that implements the ReAct algorithm. More details refer to https://arxiv.org/abs/2210.03629.

This is an example implementation of ReAct algorithm in AgentScope. We follow the idea within the paper, but the detailed prompt engineering maybe different. Developers are encouraged to modify the prompt to fit their own needs.

Note

1. We use the “thought” field in the response to support Chain-of-Thought, which means the tool functions cannot use “thought” as their argument name. 2. The function name “finish” is also a reserved name when using this agent, which will be used to end the reasoning-acting loop.

__init__(name: str, model_config_name: str, service_toolkit: ServiceToolkit, sys_prompt: str = "You're a helpful assistant named {name}.", max_iters: int = 10, verbose: bool = True) → None[source]

Initialize the ReAct agent with the given name, model config name and tools.

Parameters:

  • name (str) – The name of the agent.

  • sys_prompt (str) – The system prompt of the agent.

  • model_config_name (str) – The name of the model config, which is used to load model from configuration.

  • service_toolkit (ServiceToolkit) – A ServiceToolkit object that contains the tool functions.

  • max_iters (int, defaults to 10) – The maximum number of iterations of the reasoning-acting loops.

  • verbose (bool, defaults to True) – Whether to print the detailed information during reasoning and acting steps. If False, only the content in speak field will be print out.

reply(x: Sequence[Msg] | Msg | None = None) → Msg[source]

The reply method of the agent.

static finish(response: str) → ServiceResponse[source]

Finish reasoning and generate a response to the user.

Note

The function won’t be executed, actually.

Parameters:

response (str) – The response to the user.

to_dist(host: str = 'localhost', port: int | None = None, max_pool_size: int = 8192, max_expire_time: int = 7200, max_timeout_seconds: int = 5, local_mode: bool = True, retry_strategy: ~agentscope.rpc.retry_strategy.RetryBase = <agentscope.rpc.retry_strategy.RetryFixedTimes object>) → Any

Convert current object into its distributed version.

Parameters:

  • host (str, defaults to “localhost”) – Hostname of the rpc agent server.

  • port (int, defaults to None) – Port of the rpc agent server.

  • max_pool_size (int, defaults to 8192) – Only takes effect when host and port are not filled in. The max number of agent reply messages that the started agent server can accommodate. Note that the oldest message will be deleted after exceeding the pool size.

  • max_expire_time (int, defaults to 7200) – Only takes effect when host and port are not filled in. Maximum time for reply messages to be cached in the launched agent server. Note that expired messages will be deleted.

  • max_timeout_seconds (int, defaults to 5) – Max timeout seconds for the rpc call.

  • local_mode (bool, defaults to True) – Only takes effect when host and port are not filled in. Whether the started agent server only listens to local requests.

  • retry_strategy (RetryBase, defaults to _DEAFULT_RETRY_STRATEGY) – The retry strategy for the async rpc call.

Returns:

the wrapped agent instance with distributed functionality

Return type:

RpcObject