Why our AI operating system kept stopping too early
Our operating loop could research and choose useful work, yet still close after producing a plan. We rebuilt it around execution, slot refill and evidence-based closure.
What we tried→What happened→What we learned→Decision
The question behind the work
The problem
The system was good at discovering and describing work, but a separate task, an Owner dependency or a pending review could become a convenient stopping point. That produced sound decisions without enough material progress.
01
What we tried
- Rebuilt the operating model as one explicit intake-to-execution state machine.
- Made three outcome slots concurrent capacity rather than a daily work quota.
- Required a freed slot to pull the next highest-value executable item immediately.
- Separated real Owner gates so one blocked outcome could not freeze unrelated work.
- Treated an available independent review as executable work, not as completed administration.
02
What happened
- Adversarial tests found several ways the earlier model could report success while work remained executable.
- The rebuilt validator now derives closure from the actual queue and active slots.
- A later run was forced to resume when its next action was independent review, then continued through corrections, release and refill.
- The system still stops at genuine authority, privacy, payment and production boundaries.
03
What we learned
Autonomy is not the number of decisions an AI makes. It is the ability to carry authorised work through execution and verification.
A blocker should be attached to one outcome, not inherited by the whole portfolio.
A report is not evidence of completion; an empty executable queue is.
04 · Outcome
The decision
Keep the file-based control plane, but make work pulling, review availability and closure evidence deterministic and fail-closed.
Where the evidence leads
What this means next
Measure whether the new model produces more completed, useful outcomes without increasing unsafe actions or activity theatre.