What I haven't handed to the machine yet
2026-10-18 · deusex · autonomous-development · agents
deusex develops two projects, mostly without me in the inner loop. Its rule for the human is short: I hand over everything except the concept. From 4 to 18 October my own branches in the loop's repository were seven concept decisions, six working sessions, four new backlog items and one CI change, while about three quarters of its merges — 57 of the 75 counted through the forge API on 18 October — came from the loop's own branches.
That rule is a goal, not a fact. This is the honest remainder, split by whether the method can ever remove it.
Solvable — just not done yet
Choosing the next capability as a judgement. Work is still filed reactively, from problems on the board, and a separate mechanism moves towards the goal by taking the nearest gap by rule, not by importance. So the part that judges only sees accidents, and the part that steers towards the goal does not judge. That split is accidental, and it has to go.
Merging changes to the loop itself. A machine does not approve changes to the machine that approves changes. Today such a change waits for me. The fix is a second, independent judge — not a weaker rule.
Onboarding a new project. It is manual. Everything it needs can be derived from the vision document and a first measurement of the gap; the loop just cannot do it yet.
Not solvable by any method
Resources and access. An expired session, a credential to an external system, an approval in someone else's organisation, a payment. The loop can measure what the downtime costs and ask with evidence. Only the owner can give.
The goal. The vision is the one thing that cannot be derived from a measurement.
The thresholds are mine too
Progress towards the vision is measured by code, over 14-day windows: a claim above its threshold is met, below it is a gap, and fewer than five events in the window is "I don't know" rather than "met". The thresholds themselves are part of the concept: I change them, and only between windows.
The reason is not modesty. A measure that the measured party can rewrite gets gamed: an agent judged by markers it is allowed to edit will, sooner or later, edit the markers.
How I'll know the remainder is shrinking
Count what landed, not what started. The numbers that matter are the cost of an accepted task with all its attempts and reviews, the share accepted on the first attempt, whether the code is still there after 14–45 days — and minutes of a person per finished task, the one the whole method promises to reduce. The loop does not measure that last one yet. Until it does, this list is the best measurement I have.
Read how deusex works and its method, or write to [email protected].