LangChain agent

LangChain-based agent implementation for coshop.

Extends :class:~coshop.utils.agent.Agent with a LangChain ReAct executor, token/runtime tracking, and helper methods for copying agent state to temporary agents (used during final item ranking).

LangchainAgent(actions=[], model_name='gpt-5-nano', model_kwargs={'reasoning_effort': 'minimal'}, max_react_steps=25, verbosity=0, **kwargs)

Bases: Agent

Base class for agents that use LangChain models. Provides common LangChain functionality.

Initialize the LangChain agent.

Parameters:

Name	Type	Description	Default
actions	List[StructuredTool]	List of tools/actions available to the agent (must be set before calling this)	[]
model_name	str	Name of the model to use	'gpt-5-nano'
model_kwargs	dict	Additional keyword arguments for the model	{'reasoning_effort': 'minimal'}
max_react_steps	int	Maximum number of ReAct steps	25
verbosity	Literal[0, 1, 2]	Whether to print verbose output	0
**kwargs		Additional arguments passed to parent class	{}

cumulative_token_breakdown property

Cumulative token usage across every LLM call this agent makes — conversation turns, final-prediction / report calls, and summarizer / compression calls. Counted at the model layer (LangChainModel) and read through here; prediction/report agents share the executor's counter, so their usage is included automatically. Keys: "input", "input_cached", "output", "reasoning".

insert_message(role, content, persist_state=True)

Insert a message into the conversation history without making an LLM call.

Parameters:

Name	Type	Description	Default
role	str	Message role ("system", "user", "assistant", or "tool")	required
content	str	Message content	required
persist_state	bool	If True, appends the message to the graph state	True

reset()

Reset the agent to its initial state.

parse_langchain_response_to_actions(raw)

Parse the raw response from a LangChain model into a list of Action objects.

Parameters:

Name	Type	Description	Default
raw	List[Any]	List of BaseMessage objects from LangChain	required

Returns:

Type	Description
List[Action]	List of Action objects with content, tool_calls, and status