Services

Swarm Verify

Find duplicate work, conflicting ownership, budget overruns and incomplete shutdowns in a large agent fleet before they reach production.

For engineering teams and agent platform companies whose 100 to 500 agents share tasks, resources or spending limits.

Workflow

The first workflow Swarm Verify covers, in five stages.

  1. Claim

    An agent takes a task and holds it under a lease.

  2. Reserve

    It sets aside its share of a spending limit.

  3. Execute

    It does the work, including actions outside your systems, such as an email.

  4. Commit

    It records the result.

  5. Release

    It gives back the task, the budget and anything else it held.

Workflow

Failures

  • An agent’s lease expires, a second agent takes the task over, and the first returns late and overwrites the new result.
  • A retry repeats a step that already finished.
  • An action outside your systems, such as an email, happens twice.
  • A canceled task keeps running, or stops without giving back what it held.

Methods

Tested: we have run it ourselves. Planned: part of the design, not yet run.

  1. Model checking

    Tested

    We write the rules your agents follow as a model with stated bounds, such as how many agents, tasks and retries, and check every state it can reach within them with Matt Carey’s Everystate.

  2. Swarm experiments

    Tested

    We run the fleet with faults injected on purpose: crashes, delays, retries, leases that expire and cancellation messages that never arrive.

  3. Execution conformance

    Tested

    We check runs your fleet has recorded against the same stated rules, step by step, with Matt Carey’s Everystate.

Methods

Demonstration

Lease, budget and cancellation models, each with its unsafe and repaired design, checked with Matt Carey’s Everystate and cross-checked with TLC, and a seeded run of 128 simulated workers checked against them.

You receive

Then one loop: change the harness, run the checks, reproduce a failure, repair it and run the checks again.

  • The first transition that breaks a rule, with every step that led to it.
  • The agents it affects.
  • A replay a coding agent can use to propose a fix.
  • Regression checks that rerun before each release.

Limits

  • Every report states its bounds, its assumptions and any telemetry it was missing.
  • A small model that passes is not proof that the whole fleet is correct.
  • The public demonstration checks models of up to three agents and three tasks, and a simulated run of 128 workers. No customer fleet has been checked yet.
  • Enforcement belongs in your own queue, budget or tool service, never in Swarm Verify.

Tools

Open-source libraries for your project, each named with its maker.

Tested

Matt Carey’s Everystate
Explores every state the model can reach, and compares recorded runs with it. A TLA+ model checker written in Rust, under the MIT license.

Start

Contact

Tell us how your agents share their work. Swarm Verify is open to design partners with shared-state agent workloads.

Inquiry

A few sentences on the work you want to improve.