Retrace

The execution replay engine for AI agents.Record,replay, fork & share every LLM call your agent makes.
Open source SDKs for Python & TypeScript.

[Retrace @Retraceai_tech]

@agentmail

[Retrace @Retraceai_tech]

https://t.co/QoGbK0jCY7

[Retrace @Retraceai_tech]

https://t.co/gqHulnMFUZ scores 71/100 for AI-agent readiness. See what agents can discover, access, and use.

https://t.co/0fFukpCfzf

[Vercel Developers @vercel_dev]

Introducing https://t.co/8MeR16MRQn, a tool to measure how well agents can read your site. Backed by @oradotai's research, you can run:

▪︎ Audits with 100+ checks
▪︎ Visualizations of agents using your site
▪︎ One-click prompts to fix problems
▪︎ A CLI for agents

[Retrace @Retraceai_tech]

Debugging AI agents by replaying the exact run, not the logs

https://t.co/fLGsDnAxdC

[Retrace @Retraceai_tech]

https://t.co/oPSfoyfRZD

[Retrace @Retraceai_tech]

agents fail in ways you can't reproduce. so i built Retrace: record every run, replay it step by step, and fork from where it broke to prove a fix.

https://t.co/QoGbK0jCY7

[Retrace @Retraceai_tech]

Retrace is live on Product Hunt. Replay and fork your AI agent runs.

https://t.co/ykC37Woik5

[Retrace @Retraceai_tech]

Retrace is live on Product Hunt. Replay and fork your AI agent runs.

https://t.co/ykC37Woik5

[Retrace @Retraceai_tech]

agents fail in ways you can't reproduce. so i built Retrace: record every run, replay it step by step, and fork from where it broke to prove a fix.

https://t.co/QoGbK0jCY7

[Retrace @Retraceai_tech]

Git branching for AI agent execution.

One decorator records every LLM call. Fork from any span → your entire agent re-executes with modified input.

Record. Fork. Replay. Share.

https://t.co/QoGbK0jCY7

[Retrace @Retraceai_tech]

your agent works perfectly in the demo.then it breaks in prod and all you've got is 4,000 lines of logs and no clue which step actually went wrong.

i got tired of debugging agents like it's 2010, so i built Retrace.

think git + a video player, for AI agents 🧵

[Retrace @Retraceai_tech]

logs tell you what happened. they don't let you reproduce it.

you can't replay the run. you can't test a fix without burning tokens and praying. and you definitely can't prove the fix worked.

so most "agent debugging" is vibes and print statements.

[Retrace @Retraceai_tech]

Retrace records the whole run automatically. one wrapper around your agent function, and every model call, tool call, retrieval and error gets captured as a step.

open the dashboard and the run is a timeline you scrub through like a video. play, pause, step.

[Retrace @Retraceai_tech]

here's the part i'm actually proud of.

found the step where it broke? fork it. change the input.

Retrace replays everything before that point from the recording (basically free), then re-runs the agent live from there. the whole downstream path diverges, not just one call.

[Retrace @Retraceai_tech]

then it puts the original run and the forked run side by side, down to the first step where they split, with a verdict: improved, regressed, or unchanged.

"find the step, change one thing, prove it" beats guess-and-redeploy.

[Retrace @Retraceai_tech]

a bit on the SDK since people always ask:

• python + typescript, both first-class
• one wrapper, providers auto-instrumented (openai / anthropic / gemini)
• streams over websocket, falls back to http, buffers offline so you never lose a run

[Retrace @Retraceai_tech]

• speaks open telemetry too

https://t.co/QEfBKc9iez