TraceLog & Replay¶

Every run executed by antcrew-engine produces a TraceLog — a structured, append-only log of every LLM call, tool invocation, and intermediate result.

What gets logged¶

graph TD
    A[Agent.run called] --> B[TraceLog: run_start]
    B --> C[LLM request]
    C --> D[TraceLog: llm_request + tokens]
    D --> E{Tool call?}
    E -->|yes| F[TraceLog: tool_call + result]
    F --> C
    E -->|no| G[Parse output]
    G --> H{Valid?}
    H -->|no| I[TraceLog: validation_error]
    I --> C
    H -->|yes| J[TraceLog: run_complete]

Replay¶

Any past run can be replayed exactly from its TraceLog without making new LLM calls:

from antcrew import replay_run

# Replay a past run from its ID
result = replay_run("run_01HXYZ...")

Useful for:

Debugging — step through exactly what the agent did
Testing — assert on deterministic outputs without API costs
Audit — demonstrate to stakeholders what the model decided and why

Shipping logs to the Platform¶

from antcrew import Agent
from antcrew.platform import PlatformSink

agent = Agent(
    model="openai:gpt-4o-mini",
    trace_sink=PlatformSink(
        base_url="https://platform-int.antcrew.org",
        api_key="your-key",
    ),
)

All events are streamed to the platform in real time and visible in the Runs dashboard.