Quiesce

A deterministic shutdown-assurance harness that proves autonomous agents have truly stopped—across descendants, credentials, jobs, retries, queues, and post-STOP effects.

Comment

Inspiration

Modern AI systems can stop responding while still leaving active child agents, valid credentials, queued work, retries, or delayed side effects behind. Existing evaluations focus on capability, not whether an autonomous system has actually stopped. I wanted to build a deterministic way to prove shutdown instead of assuming it.

What it does

Quiesce is a deterministic shutdown-assurance harness for autonomous agents.

It simulates shutdown scenarios, injects STOP, measures residual authority, pending work and escaped effects, and determines whether the system truly reached quiescence. Every PASS or FAIL is produced by deterministic verification rather than by an LLM.

How we built it

Quiesce combines a deterministic shutdown engine, protected replay, authority epochs, cascade revocation, commit fences, a Quiescence Sweep, tamper-evident certificates, and bounded GPT-5.6 roles for contract compilation and incident narration. GPT-5.6 never decides whether a test passes or fails.

Challenges we ran into

The hardest challenge was separating deterministic verification from language-model reasoning while still using GPT-5.6 where it genuinely adds value. We also designed scenarios that clearly demonstrate shutdown failures without relying on randomness.

Accomplishments that we're proud of

  • Deterministic shutdown verification.
  • Proof that normal STOP can still leave residual authority.
  • Protected replay reaching verified quiescence.
  • Tamper-evident shutdown certificates.
  • A clear visual demonstration of shutdown assurance.

What we learned

Reliable AI systems need verifiable shutdown guarantees, not just graceful exits. Combining deterministic verification with narrowly scoped language-model assistance creates systems that are easier to trust and audit.

What's next for Quiesce

Expand the scenario library, integrate with real multi-agent platforms, support continuous shutdown testing in CI/CD pipelines, and evolve Quiesce into a standard verification layer for autonomous AI systems.

Built With

Share this project:

Updates