Deep Agents

Auto-trace LangChain Deep Agents with enriched visibility into planning, subagent delegation, filesystem operations, and sandbox execution.

Installation

pip install "axonpush[deepagents]"

[!NOTE] Requires Python 3.11+ (the deepagents package does not support 3.10).

Setup

import os
from deepagents import create_deep_agent
from axonpush import AxonPush
from axonpush.integrations.deepagents import AxonPushDeepAgentHandler

client = AxonPush(
    api_key=os.environ["AXONPUSH_API_KEY"],
    tenant_id=os.environ["AXONPUSH_TENANT_ID"],
    base_url=os.environ.get("AXONPUSH_BASE_URL", "https://api.axonpush.xyz"),
)

handler = AxonPushDeepAgentHandler(
    client=client,
    channel_id=1,
    agent_id="deep-agent",
)

Usage

Pass the handler via config to any Deep Agent invocation:

agent = create_deep_agent(
    tools=[my_tool],
    system_prompt="You are a helpful assistant.",
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Research AI frameworks"}]},
    config={"callbacks": [handler]},
)

Events Traced

The handler automatically classifies Deep Agent built-in tools into enriched event identifiers:

Event	When
chain.start	Agent chain begins execution
chain.end	Agent chain completes
chain.error	Agent chain errors
llm.start	LLM call begins
llm.end	LLM call completes
llm.token	Streaming token received
planning.update	write_todos tool called (planning step)
planning.complete	write_todos tool completes
subagent.spawn	task tool spawns a subagent
subagent.complete	Subagent finishes
filesystem.read	read_file, ls, glob, or grep called
filesystem.write	write_file or edit_file called
sandbox.execute	execute tool runs a shell command
tool.<name>.start	Any other tool begins
tool.end	Any other tool completes
tool.error	Tool execution errors

All events include framework: "deepagents" in metadata.