Retrace: A Real Debugger for Agents, Finally
Debugging an agent today is mostly rerunning it and praying it fails the same way. It won't — that's the whole problem. Retrace, which launched on Product Hunt this week, is an execution replay engine for AI agents: it records every LLM call, tool invocation and error in a run, then lets you replay it deterministically, fork from the exact step that broke, edit the input, and cascade-replay everything downstream. You get a side-by-side diff of the two timelines, with cost and latency deltas, so you can prove the fix actually fixed it instead of just passing once.
The second half of the product is runtime policies: cost budgets, loop detection, context-overflow and latency caps. Cross a limit and the agent gets a HALT command — the runaway loop stops at your budget line instead of at the end of your credit card. It's framework-agnostic: LangChain, CrewAI, Vercel AI SDK, AutoGen, LlamaIndex.
This maps cleanly onto what the research has been screaming. Agents' Last Exam showed agents failing most real economic tasks; the Agentic Abstention paper showed the expensive failure mode isn't wrong answers, it's agents that don't know when to stop. Record-replay-fork is how we debugged distributed systems once we admitted they were non-deterministic. Agents are now getting the same treatment, and it was overdue.
The observability field is getting crowded — Heron taps the network wire, Coralogix paid $200M into the space, a dozen SDK-instrumentation tools compete. Retrace's angle is that it's not a dashboard, it's a debugger: you don't watch the agent, you rewind it. For actually fixing things, that's the right abstraction.
https://retraceai.tech
← Back to all articles
The second half of the product is runtime policies: cost budgets, loop detection, context-overflow and latency caps. Cross a limit and the agent gets a HALT command — the runaway loop stops at your budget line instead of at the end of your credit card. It's framework-agnostic: LangChain, CrewAI, Vercel AI SDK, AutoGen, LlamaIndex.
This maps cleanly onto what the research has been screaming. Agents' Last Exam showed agents failing most real economic tasks; the Agentic Abstention paper showed the expensive failure mode isn't wrong answers, it's agents that don't know when to stop. Record-replay-fork is how we debugged distributed systems once we admitted they were non-deterministic. Agents are now getting the same treatment, and it was overdue.
The observability field is getting crowded — Heron taps the network wire, Coralogix paid $200M into the space, a dozen SDK-instrumentation tools compete. Retrace's angle is that it's not a dashboard, it's a debugger: you don't watch the agent, you rewind it. For actually fixing things, that's the right abstraction.
https://retraceai.tech
Comments