Retrace

About Retrace

Retrace is a debugging tool that records, replays, and forks AI agent executions. It captures every LLM call, tool invocation, and error as a span inside a trace, then lets you play back the run step by step. The tool is free for up to 1,000 traces per month.

Review

Debugging an agent that fails at step eight of a twelve-step chain usually means re-running the whole thing just to test a single fix. Retrace records the full trace and replays it like a scrubbable video, so you can find the exact break point, fork from there, and let only the downstream LLM calls re-execute live. It treats forking more like a git branch than a prompt re-run - pre-fork steps stay on tape, and tool outputs are replayed instead of re-fired.

Key Features

• Record and replay traces. Every LLM call, tool call, and error appears as a span in a nested, scrubbable timeline. You replay the run step by step to see what happened.
• Fork with live re-execution. Pick any span as a fork point. Steps before the fork replay from the recording; downstream LLM calls run live, while tool calls replay their recorded outputs to avoid side effects.
• Output override and mock tagging. Edit the recorded output at any overridden span. The edited value gets tagged as "mocked" and flows into the live downstream steps. A per-step map (span id → output) supports scripted sweeps of different "what-if" values.
• First-divergence diff and verdict. After a fork, the tool highlights the first step where the new path diverges from the original and classifies the result as improved, regressed, or unchanged.
• Multi-agent topology view. Spans carry agent IDs and roles. An agent-topology graph shows how sub-agents hand off work, with inter-agent detectors flagging reasoning/action mismatches.

Pricing and Value

Retrace currently operates on a free tier that includes 1,000 traces per month. No other pricing plans have been published yet.

Pros

• Tool calls are not re-executed during a fork, so debugging production runs with payments, emails, or checkout flows doesn't trigger duplicate side effects.
• The fork-and-replay model allows testing a fix at a specific step without re-running the entire agent chain from scratch.
• Override maps let you script "what if" scenarios by supplying alternative tool outputs without touching live APIs.
• Traces display original and forked paths side by side, which helps isolate exactly where a change altered behavior.
• The replay timeline groups nested calls, so multi-step workflows stay readable instead of becoming a flat list of events.

Cons

• Replay matches tool calls positionally rather than by name or arguments. When a forked model calls the same tool with different parameters, it gets the old taped answer, and a new tool call has no tape entry at all. This makes the tool less authoritative for agents that frequently change tool call arguments.
• Handling very large traces with dozens of tool calls and multiple sub-agents is still being actively hardened, so performance and clarity may drop on unusually complex runs.
• Teams whose debugging workflow requires actual live re-execution of side-effecting tool calls with fresh arguments will find the tape replay model insufficient. The tool is not a fit when real-world API side effects must fire during a fork.

Retrace works best for developers building agents that need to inspect prompt chains and non-deterministic model behavior without constantly re-running expensive or side-effecting workflows. It's particularly useful for catching looping tool calls, testing prompt or model changes mid-run, and performing regression checks with override values. Teams that can work within the positional replay limitation and don't need live side-effect re-execution will get the most out of its current design.

Open 'Retrace' Website