Skip to content

autogen.md

Latest commit

 

History

History
130 lines (95 loc) · 4.31 KB

autogen.md

File metadata and controls

130 lines (95 loc) · 4.31 KB
Raw

AutoGen / AG2

Trace AutoGen multi-agent message passing, LLM calls, and tool invocations with a callback handler that captures agent-to-agent communication.

Installation

pip install lantern-ai[autogen]

This installs lantern-ai along with autogen-agentchat>=0.4.0 as a dependency.

Setup

from lantern_ai import LanternTracer
from lantern_ai.collectors.autogen import create_lantern_autogen_handler

tracer = LanternTracer(
    api_key="your-lantern-api-key",
    base_url="https://your-lantern-instance.com",
)

handler = create_lantern_autogen_handler(tracer, agent_name="my-autogen-team")
# agent_name is optional, defaults to "autogen-agent"

Usage

Trace agent-to-agent messages

AutoGen's multi-agent conversations involve agents sending messages to each other. Track each message exchange:

from autogen import AssistantAgent, UserProxyAgent

assistant = AssistantAgent("assistant", llm_config={"model": "gpt-4"})
user_proxy = UserProxyAgent("user_proxy", code_execution_config=False)

# Track a message from user_proxy to assistant
span_id = handler.on_message(
    sender="user_proxy",
    recipient="assistant",
    content="Write a Python function to calculate fibonacci numbers",
)

# After the message is processed:
handler.on_message_end(span_id=span_id, output="Here's the fibonacci function...")

Trace LLM calls

span_id = handler.on_llm_start(
    model="gpt-4",
    messages=[{"role": "user", "content": "Write fibonacci in Python"}],
)

handler.on_llm_end(
    span_id=span_id,
    output="def fibonacci(n): ...",
    input_tokens=15,
    output_tokens=89,
)

Trace tool calls

span_id = handler.on_tool_start(tool_name="execute_code", args={"code": "print(fibonacci(10))"})
handler.on_tool_end(span_id=span_id, output="55")

Full conversation example

# Start tracking
handler.on_message(sender="user_proxy", recipient="assistant", content="Solve this math problem: 2+2")
span_id = handler.on_llm_start(model="gpt-4", messages=[{"role": "user", "content": "2+2"}])
handler.on_llm_end(span_id=span_id, output="4", input_tokens=5, output_tokens=1)
handler.on_message_end(span_id=msg_span_id, output="The answer is 4")

handler.finish()

What Gets Traced

AutoGen Event Lantern Span Type Captured Data
on_message / on_message_end custom Sender, recipient, message content, response
on_llm_start / on_llm_end llm_call Model name, messages, response, token counts
on_tool_start / on_tool_end tool_call Tool name, arguments, output

Message tracking

Agent-to-agent messages are tracked with a composite key (msg_{sender}_{recipient}). Each message exchange becomes a custom span showing the communication flow between agents.

Troubleshooting

Spans not appearing

  • Make sure finish() is called when the conversation ends.
  • Verify api_key and base_url are correct.

Message spans overlapping

  • Messages between the same sender-recipient pair overwrite each other in the active spans map. If agent A sends two messages to agent B without the first one ending, the second will replace the first.

Token counts always zero

  • Token counts must be passed explicitly to on_llm_end(). AutoGen doesn't always expose token usage from the underlying LLM.

API Reference

def create_lantern_autogen_handler(
    tracer: LanternTracer,
    *,
    agent_name: str | None = None,
) -> LanternAutoGenHandler

LanternAutoGenHandler methods

Method Parameters Returns Description
on_message sender: str = "", recipient: str = "", content: str = "" str (span_id) Start a message exchange span
on_message_end span_id: str, output: str = "" None End a message exchange span
on_llm_start model: str = "", messages: list = None str (span_id) Start an LLM call span
on_llm_end span_id: str, output: str = "", input_tokens: int = 0, output_tokens: int = 0 None End an LLM call span
on_tool_start tool_name: str, args: Any = None str (span_id) Start a tool call span
on_tool_end span_id: str, output: str = "" None End a tool call span
finish — None End the trace