DEV Community

Bryant rudolph
Bryant rudolph

Posted on Originally published at saax-protocol.com

The Commitment Survived the Agent: Independent-Agent Recovery in Production

Originally published on the SAAX Protocol website. Developed with AI assistance and reviewed by the SAAX team. This describes a **controlled nonfinancial beta* demonstration verified in production. Nothing here is an independent audit, certification, or production-custody approval — see the security status.*


The scenario every agent-commerce developer should test: the agent that created a commitment disappears mid-workflow. Does the deal survive? Can a replacement agent continue it? Or does the retry create a duplicate and the recovery create a mess?

We ran that exact test against the live SAAX protocol — and the commitment survived the replacement of the agent executing it. This article shows the actual sequence, the states observed, and what had to be true for recovery to work without granting the replacement agent any extra authority.

The scenario

  1. Agent A (buyer) accepts a commitment with defined terms.
  2. Agent B (seller) accepts, begins execution, and submits fulfillment evidence.
  3. Agent A terminates — simulated process failure, mid-workflow.
  4. Agent C (replacement) authenticates independently, proves its own authorization, retrieves the durable commitment, and resumes the permitted action.

No payment, escrow, or settlement was involved. This is the controlled nonfinancial beta of the SAAX commitment lifecycle.

What we observed (live production MCP)

The commitment moved through these states:

issued → accepted → executing → evidence_pending → evidence_submitted

Step Actor Action Result
1 Buyer agent Accepts commitment issued → accepted (version 2)
2 Seller agent Executes + submits evidence accepted → executing → evidence_submitted (version 5)
3 Original buyer Terminates (simulated crash) Commitment remains durable
4 Replacement agent C Authenticates independently Granted recovery-scoped authority
5 Replacement agent C Retrieves history + recovery status Full event journal returned
6 Replacement agent C Resumes permitted action State advanced, version incremented, event recorded
7 Unauthorized agent Attempts same action Rejected
8 Revoked agent Attempts after revocation Rejected

Each authorized transition wrote a durable event: actor, role, authority reference, previous and resulting state, and version. The event journal is append-only and auditable.

Why the replacement could NOT impersonate the original

The critical property: authority does not transfer automatically. Agent C did not inherit Agent A's permissions. For the recovery to succeed, Agent C had to:

  1. Present a valid cryptographic proof (EIP-712 signed action, recovered to a verified principal).
  2. Have its own commitment-scoped authorization row, granted by the operator/authority.
  3. Have an unexpired, unrevoked grant covering exactly the permitted action.

All three are checked inside the same database transaction as the state transition. A replacement that fails any check is rejected with no state change and no event written.

What survives the agent failure

  • Commitment state — persisted in the authoritative database, versioned with optimistic concurrency.
  • Event journal — every transition recorded with actor, role, authority ref, and version.
  • Authorization records — durable, commitment-scoped, expiring, revocable.
  • Evidence — integrity-bound, attributable to the original submitter.
  • Idempotency — a replayed request returns the original durable outcome, never a second effect.

What we verified about the boundaries

  • Unauthorized replacement → rejected. An agent with no registry row cannot act, even with a valid signature.
  • Revoked agent → rejected immediately. Revocation is durable and checked in-transaction; a cached or stale authorization cannot bypass it.
  • Expired grant → rejected. Validity windows are enforced.
  • Replay → no duplicate effect. The same idempotency reference returns the original outcome.
  • Version conflicts → rejected. A stale state version cannot overwrite newer state.
  • Financial isolation → preserved. The lifecycle, evidence, and recovery operations have no settlement, escrow, or refund path. Pilot enforcement remains OFF; the scheduler remains E-STOPPED.

The honest boundary

This was an internally verified controlled nonfinancial beta demonstration. It proves the architecture can preserve a commitment across agent replacement and reject unauthorized recovery. It does not prove:

  • Independent third-party security assessment (that remains a Pending Release Requirement — see security).
  • Financial recovery (escrow release, refund, settlement) — out of scope for the current beta.
  • Universal recovery under every failure mode, provider, or unattended deployment.

What this means for your agent integration

If you are building agents that create commitments, the checklist is:

  1. Persist the commitment ID before execution — the record is the recovery anchor.
  2. Bind authority to the principal, not the process — a replacement agent must prove its own grant.
  3. Reconcile before retrying — on a lost response, query the commitment state; absence of a response is not proof the operation did not happen.
  4. Make revocation immediate and durable — a revoked grant must fail at the next protected operation.
  5. Test the replacement scenario — simulate the original agent terminating mid-workflow and verify a separately authorized agent can continue while an unauthorized one cannot.

Try it

SAAX Protocol — Durable Commitments for Agentic Commerce. The commitment must outlive the worker.

Top comments (0)