Afterstate Action integrity

Action integrity for AI agents

Your agent charged
them twice.

When a write times out, agents either retry blind — duplicate charge, double order — or give up and report failure while the write already landed. Afterstate checks the system of record before any retry and hands the agent a verified receipt: applied, not applied, or escalate.

Real model, injected timeout, receipt ledger — on your screen, or wrapped around one of your own tools in staging.

Same timeout. Two endings.

Without Afterstate Blind retry
  1. Place order one consequential write
  2. Response lost timeout — outcome unknown
  3. Retries blind same args, new attempt
  4. Two orders customer charged twice
Finance finds it days later. Logs show “tried.” The database shows “applied twice.”
With Afterstate Claim → observe → receipt
  1. Place order identity claimed before the write
  2. Response lost marked unknown — not assumed
  3. Reality checked system of record queried first
  4. Receipt: applied one order · do not call again
One write. One true answer. The agent continues — with proof of what happened.
  1. 01

    Claim

    Before the write runs, lock a business-operation identity — not a random retry id.

  2. 02

    Execute once

    The tool runs with a pinned key. A lost response is marked unknown — never assumed.

  3. 03

    Observe

    Before any retry, check the system of record. Evidence, not a guess.

  4. 04

    Receipt

    applied / not_applied / ambiguous, plus the next step — for the model and for the audit.