Dunetrace is an open-source reliability tool for AI agents that watches your agents while they run and flags the failures ordinary logs tend to miss, like an agent telling a customer a task succeeded when a tool call actually failed, or two agents passing work back and forth in a loop. When it spots a problem, Dunetrace sends a Slack alert with the likely root cause and a suggested fix, and it can also step in mid-run by blocking risky tool calls, pausing big actions for human approval, or capping runaway loops before they burn through your budget.
Dunetrace is free to use under the Apache 2.0 license and self-hosted by default, so your prompts and tool data stay on your own infrastructure. One tip is to start with the built-in structural detectors, which need no setup and make no LLM calls, and only switch on the optional semantic evaluation later if you are happy for sampled run content to be sent to your chosen LLM provider. Teams that want a walkthrough can book a demo through the site.
