Jan 2026–present / Project
Making agent work accountable.
AI Project System
The context
An agent can report success without producing useful work. AI Project System explores how to organize software delivery around explicit scope, observable actions, reviewable artifacts, and human authority.
My contribution
The delivery workflow
Developed a governed workflow for defining work, dispatching agents, checking results, and recording what happened before deciding the next step.
Boundaries and review
Built bounded execution and retry paths, with human approval for consequential decisions and safeguards when required evidence is missing.
Experiments in public
Documented cases where successful process exits or confident QA verdicts did not correspond to completed work, alongside the evidence from a completed delivery run.
What came out of it
A public system with inspectable delivery records and documented failure cases. The completed Epic is evidence of one execution route; it is not a claim that arbitrary development tasks can run reliably without oversight.