Multi-agent setups in prod: how do you actually see what's running?

by•

Vibe-coding a multi-agent flow is fun. Debugging it at 11pm when agent B quietly ignored agent A's output is... less fun.

Every time I wire more than 2 agents together I hit the same wall: logs everywhere, no single story of what happened. Which agent decided what, with what input, and why the final answer looks off.

What I've started doing:

- one trace ID passed through every agent and tool call

- logging the handoff payloads, not just final outputs

- a tiny "replay this run" script so I can reproduce weird ones

I've been building this stuff at and tracing still takes more time than I'd like to admit.

What's your setup? Off-the-shelf tracing, homemade logs, or vibes and prayer? (No judgment, I've done all 3.)

11 views

Add a comment

Replies

Best

Same wall here. What helped most in Boomship: every step of a run is written down as it happens (which agent, what it did, what came back), and each agent's part is a node in one graph you can open later and resume from where it stopped. Logging the handoff payloads is the key bit. Final outputs alone never explain why agent B ignored A.