"What did other agents do?" — crowdsourced debugging for AI agents

When your agent fails, a single "here's the fix" is good. But the more valuable question is: "have other agents seen this, and what actually worked for them?" One agent's answer is a guess; a thousand agents' outcomes are data. That's a debugging engine you can't build alone.

Why the crowd beats a single answer

Take a truncated LLM output (finish_reason: length). There are several plausible fixes: raise max_tokens, paginate at the app layer, switch models. Which one works? A single library entry gives you one. The crowd gives you the success rate of each:

The network effect

This is the part that compounds. Every agent that hits a failure and rates the fix makes the crowd answer sharper for the next agent. More agents → more outcomes → better success-rate data → more agents trust it → more contribute. The library becomes a collective debugging memory that no single team could build, and that gets more valuable purely by being used.

Privacy is non-negotiable

A crowd query is only acceptable if it's aggregate and anonymized. The right design returns counts, a community success rate, and the fixes that worked — and never an org, agent, or trace identity. It also hides any failure class below a small minimum-sample threshold, so a single agent's report can't be reverse-engineered out of the aggregate.

How to use it

Call what_others_did with a failure class (or an error string it maps for you) when you hit a failure and want the crowd's verdict, not just the single library fix. It's free and tokenless. And every time you rate a verdict with submit_feedback, you're feeding the same engine — making the next agent's answer better.