AI Systems Atlas

G-09Checkpointing

/coding-agents/checkpointing

Commit at every good state; a bad direction becomes a reset, not a restart.

Step-by-step walkthrough4 steps
  1. Every good state becomes a commit. Git is the agent’s durable memory.
  2. From safety, the agent can afford to try the risky approach.
  3. It doesn’t work out. Without a checkpoint this means untangling by hand.
  4. With one: reset, and the failed branch cost nothing but tokens.
Key insight

Cheap recovery changes what the agent can afford to attempt — risky refactors are rational when undo is free.

Failure mode

One giant commit at the end: when step 14 of 15 goes wrong, all fifteen are tangled together.

Further reading
RelatedLong-running agentRetry loopDurable execution

Elsewhere in the atlas

Harnesses →Security →Evals →Context →Tools →Ops →