What is the evaluate-then-commit pattern for AI agents?

The evaluate-then-commit pattern separates an AI agent's action proposal from its execution, creating a deterministic checkpoint where every proposed action is validated against policy rules before any state-changing operation occurs.

In standard agent frameworks, reasoning and execution are coupled:

LLM decides → Tool executes → State changes (no validation)

The evaluate-then-commit pattern decouples them:

LLM decides → Exogram evaluates → Policy passes? → Tool executes → State changes

This is Exogram's EAAP protocol (Exogram Action Admissibility Protocol):

1. /v2/evaluate: The agent submits a proposed action. Exogram evaluates it against 8 policy gates in 0.07ms. Returns an execution token (JWT) if approved.
2. /v2/commit: The agent submits the execution token to actually perform the action. Exogram verifies the token hasn't expired, the state hasn't drifted (SHA-256 hash comparison), and the action matches the evaluation.

If anything changes between evaluate and commit — state drift, token expiry, concurrent modification — the commit is rejected with 409 Conflict. The agent must re-evaluate. This is the Execution Idempotency invariant: no action commits against stale state.