Tool Trail

Choose what an agent should do when a tool call fails.

Choose Retry, Roll back, or Stop after each tool event while keeping external state trustworthy.

How to play

Four tool events interrupt an agent’s run. Choose retry, rollback, or stop and inspect. A wrong choice loses trajectory integrity. Read the outcome after each decision to see what happened to external state.

The systems concept

A retry is safe when the operation is read-only or idempotent. An effectful call may have succeeded even if its reply was lost. Compensating actions can undo some effects, while uncertain or unverified state may require stopping before another action. This game reduces those recovery choices to four short cases.

Explore Volume III: Agentic Machine Learning Systems or return to the playground.