New open sourced eval tool for building agents

**agent_detective** is a new open-source evaluation tool from GitHub user **Thomeras**, shared under **Thomeras/agent_detective**. The author posted it on Hacker News (item id **49152073**) asking which agent broke in a multi-agent system or workflow, and invited the community to try it. At the time of the post it had **2** points and **1** comment.

The tool’s stated job is narrow and operational: given a system or workflow made of multiple agents, answer **which of the agents broke**. That framing treats multi-agent failure as a localization problem rather than a vague “something failed” outcome. As an eval tool for building agents, it sits on the path of running a workflow, observing failure, and attributing blame to a specific agent instead of the whole graph.

For engineers building multi-agent systems, that attribution is the expensive part of debugging. When several agents chain tools, share state, or hand off tasks, a single bad step can surface as a downstream error. A tool that targets **which agent broke** reduces time spent replaying full workflows and guessing which hop introduced the fault. Builders shipping agent workflows need that kind of failure isolation as much as they need better prompts or model upgrades.

On the market side, agent stacks are proliferating while open eval and debugging tooling is still thin relative to production frameworks. **agent_detective** is an early community contribution in that gap: not a full platform announcement, but a focused open-source attempt at agent-level failure detection. The thin HN engagement so far (**2** points, **1** comment) means it has not yet been stress-tested in public discussion; usefulness will depend on whether builders adopt and report real workflows against it.

Practical next steps are straightforward. Clone **https://github.com/Thomeras/agent_detective**, run it against a multi-agent workflow you already operate, and check whether it correctly names the failing agent. Watch whether the author documents supported workflow shapes, failure signals, and integration points; those details will decide if it stays a one-off experiment or becomes a reusable eval piece in agent development.