Idea Lab · Oct 19 2026 · AI idea demo

Understudy

Understudy records an agent's real runs (the prompt, every tool call and every tool result) and replays them against a candidate model, serving the recorded tool results back so nothing live is touched. It then lines the two runs up step by step and flags where the candidate decided differently: a different tool, changed arguments, an extra write, or a skipped confirmation. This page shows how a replay report would read.

Example data: the runs, models and decisions below are invented to show the report format. No real model was replayed.

What to try

• Open the "Pay vendor invoice" run and switch the candidate to Model C: the replay skips the approval step, which is the kind of change Understudy exists to catch.
• Tick "Divergences only" to see the report the way a reviewer would skim it before approving a model switch.
• Compare "Triage support inbox" across both candidates: different wording in arguments is flagged as a soft change, while an extra write is flagged as a hard one.