Get the weekly radar
One short email with the freshest products, every week.
Retrace records agent execution, lets you fork from the failing step, and proves your fix worked.
Grounded in available product and source data
An agent fails on step 4 of a 10-step run, and the usual fix is re-running the whole thing from scratch hoping the bug shows up again — Retrace is fork and replay debugging for AI agents built specifically to skip that: branch execution from the exact failing step instead of the beginning.
A single decorator is what instruments an agent, capturing every LLM call, tool call, and error and streaming it into a live, span-level trace UI as execution happens. When something breaks, Fork & Replay branches from that specific step and cascade-replays forward with the original context intact, showing a side-by-side diff of cost and latency against the first attempt. Prove the Fix takes it further — re-run a proposed change against the failed trace and get back a direct verdict: improved, regressed, or unchanged, without needing to re-execute the entire agent to find out.
Beyond debugging after the fact, runtime enforcement adds guardrails that watch agent behavior live, with circuit breakers that halt execution outright at budget, loop, or latency limits, and a pre-call gateway that can hold an action for approval before it runs. Quality gates catch subtler problems automatically — groundedness gaps, statistical drift, failure clusters — and CI/CD integration means a regression can block a deploy in GitHub Actions rather than reaching production first.
Framework coverage is broad: LangChain, CrewAI, Vercel AI SDK, AutoGen, and LlamaIndex are all supported, alongside OpenAI, Anthropic, and Google Gemini as LLM providers, with Python and TypeScript SDKs. Setup is stated to take under 2 minutes — install, add a decorator, run. Five tiers scale primarily by trace volume, retention window, and fork-replay allowance, from a Free tier capped at 1,000 traces monthly up through Enterprise with unlimited traces, replays, and users, with CI regression gates and multi-agent detectors specifically gated to Pro and above. One thing not fully available yet: bringing your own API key currently only works with Google Gemini, since OpenAI and Anthropic key support is listed as still coming.
No — Fork & Replay branches from the exact failing step and cascade-replays forward from there, so a fix gets tested against the specific failure point rather than requiring a full re-execution from the start.
Prove the Fix returns a direct verdict — improved, regressed, or unchanged — measured against the original failed run, rather than leaving a developer to eyeball the difference manually.
No — it's not purely a post-mortem tool; the enforcement layer intervenes live while an agent is still running, which is the difference between reading about a runaway loop afterward and actually stopping it mid-run.
No — it works across LangChain, CrewAI, Vercel AI SDK, AutoGen, and LlamaIndex, with the site framing this as no lock-in to one particular agent framework or provider.
Mainly volume and retention — Free caps out at 1,000 traces monthly with 7-day retention and 10 fork replays, while Starter raises those to 10,000 traces, 30-day retention, and 100 fork replays, plus adds prove-the-fix runs and Cassette VCR replay.
These 1 comments were collected from the discovery source. ProductVisit does not host replies in Release A.
Retrace records every LLM call, tool call, and error in a run as a span inside a trace. You can replay a past run step by step, like scrubbing through a video. When you find the step that broke, you fork it, change the input or model at that point and the agent re-executes from there, so you can compare the original and the new path side by side. The part I care most about is the forking: it's closer to git branching than to re-running a prompt. Pre-fork steps replay from the recording; everything downstream runs live. It's early, and I'd really like your feedback — especially on the replay and fork flow, and what would make it fit your stack. Which frameworks or providers are you using? Happy to answer anything here.