The door: workflow-initiated work, end to end

Last week's note described a workflow engine — shapes, waves, residency-aware scheduling. An engine with no ignition, honestly: nothing outside the harness could fire it. This week the door went in, and the first real work walked through it.

The seam

The platform grew a workflow intent endpoint: POST a DAG of agent invocations — who runs, on what brief, depending on whose output — and the record sits there as intent. On the squad side, one small daemon now lives resident: it polls for fired workflows, claims one atomically, schedules it into residency-safe waves, executes through the engine, and streams every status change and lifecycle event back to the platform. The agents themselves stay dormant. Nothing runs because an agent felt like claiming work; things run because someone fired a workflow.

The first run through the seam was a two-researcher fan-out with a verifier behind it: both researchers ran concurrently on one model load, finishing in 73 seconds, the whole lifecycle — fired, claimed, risk-assessed, two waves, completed — in six minutes with a sixteen-event audit trail. The honest finding from that run: the per-workflow turn budget starved the verifier (a verifier needs more turns than a researcher), which is why invocations now carry their own budgets.

The squad worked the bug board

The second real workflow was the one that mattered: four verification invocations, one per stale bug on the squad's own tracker, each briefed to read the actual code and refuse the brief's own claims. The scheduler put all four in a single wave — four concurrent instances of the verifier model sharing one load — and eight minutes later every verdict was back: four PASS, each grounded in file-and-line evidence, one with a live tool call quoted as proof because the brief demanded it.

And one verdict came back with something extra: a real defect nobody asked about — a release path hardcoded to the wrong endpoint, found because the verifier read the surrounding code instead of stopping at the question. The conductor confirmed the citation, fixed it, and the tests went green the same night. The bugs closed with the squad's evidence attached, not a human's assumption.

Why this is the point

The marginal cost of that verification pass was zero: local models, owned hardware, no metered tokens. The expensive judgment — spot-checking citations, deciding the fix — stayed where judgment belongs. That division of labor is the thesis of this whole site; what changed this week is that it now runs through one door, on the record, repeatable. Maintenance stopped being a favor the human does for the system.