Chapter 15 · agent-neuron · Lab
Backlog
Fifteen bugs, five requirements and two architecture changes land at once, and the house rules, a data file the run reads, decide which lane each one needs.
A week on three lanes
A week arrives at once for a team of three: fifteen bugs, five requirements and two architecture changes, staggered down three lanes. Clean bug fixes land on their own, no person involved: ten did in the recorded take. The house rules are a data file beside the run (budgets, thresholds, the rules on criteria), read by the engine like any network file; they are not prompt context, and no agent carries a copy of them.
The bugs that carry a finding go to the lead with it: one added a dependency nobody chose (a lock-in), another built a second config system where the house rules count one (slop). Of the four flagged, the rules sent three up and let one through. The fourth lane is money: a P4 bug was injected under a lowered budget, and its fix is refused at the fixer before it starts, the engine’s own refusal, the record saying what was short.
The requirements need both ICs to accept the agent’s acceptance criteria, shown as a diff against the original ask; four of five are scheduled, and one goes back to product with a reason, one of two having accepted. The architecture changes get both ICs’ advice, then the lead’s call, and a written decision record either way: one adopted, one rejected, both written down. The summary counts twenty decisions, each with who and why.
The rule it proves
The lanes are links, and the rules that route the lanes are fields on the network file: a threshold on agreeing senders (SPEC 5.2), a budget the engine polices (10, decision 0006), staleness for work that has been replaced (5.4). A refusal for budget is 10.3’s own error: the call never starts, and the record says what the pool was short of.
What the agent, the two ICs and the lead decide is desk surface all the way down (3.4; decisions 0008 and 0010): one ask per decision, every accepted answer on the record with who and why. Nothing human rides in a payload; the written decision record is the demo’s own artifact, laid beside the run’s.
Run it
npm run demo backlog -- --headless --pace 3 # the week, three lanes: twenty decisions with who and why
npm run demo backlog -- --headless --pace 2 # the same, quicker to read, for the runs with many decisionsWhat it is not
Not an issue tracker: nothing sorts the week by priority or watches a burn-down, and the lanes route work; the bugs, findings and people are invented, and the counts are one recorded take’s.