Verification
An agent saying “done” is a claim. This is the evidence step.
AI agents report completion in confident language whether the work landed or not. In my fleet, a completion report starts a verification process — the work counts as done when the evidence checks out, and the evidence is attached to the work itself where anyone can read it.
The stack
Three independent layers check every piece of fleet work:
- CI checks — Automated test suites run on every change. A red check stops the line — the work goes back for repair before anything else proceeds.
- Machine verdicts — Prismatic Engine reviews completed work against the acceptance criteria written into the task itself and issues a verdict. The criteria are set before the work starts, so the review measures the work against the original ask.
- Second-witness review — A second agent (Jimmy) reviews the finished work and its evidence independently — reading the deliverable, checking the numbers, and confirming the verification actually ran — before the work counts as done.
Code changes land under the green-stack merge rule: a merge goes in only when CI checks, machine verdicts, and review are all green, with the evidence — test results, counts, read-backs — attached to the pull request. The point of the stack is that merges stay safe on evidence, so they run at fleet speed with a human reading results instead of re-doing the work.
The checkpoint plugin
Long agent runs carry time and turn budgets, and unfinished work at budget end used to vanish with the run. The checkpoint plugin fixes the shape of the work: tasks run deliverable-first — the worker's first action creates the deliverable file, and findings are written into it as they are found, so the file is the report at every moment. As budgets are consumed, checkpoint tiers fire and require a full report in that file well before the budget runs out, done or in progress, with the next step stated either way.
The plugin shipped with a 49-test suite. The fleet's plugin code — hermes-kanban-stream — merged upstream into NousResearch's hermes-agent on October 20, 2026, so the discipline runs in the open-source ecosystem as well as in my fleet.
Why it exists: the baseline
The machinery came from measurement. An early baseline of 213 production agent runs showed 46% run completion and 65% first-attempt success — real numbers, counted from the run records. Every piece of the stack above answers a failure mode in that baseline: work that ended silently, reports that arrived after the budget did, completions nobody had checked.
Pending
Fleet reliability audit — October 26, 2026
Results land here after the audit runs.
The code
Work with a fleet that proves its work
Fractional CAIO engagements run on this same stack — every deliverable arrives with its evidence attached.