Concept register · Theme 06 of 14 9 concepts · 93 talks

assay  ·  concepts  ·  the-repo-is-a-factory

The repo is a factory

Continuous AI as a third pillar beside CI/CD: cadenced agent loops inside the repository, read-only sandboxed steps, narrow countable output channels, and the pull request as the hard human gate.

assay: strongest external validation

Harness engineering as a discipline — 25 sources, establishedModel tiering and routing — 21 sources, establishedMulti-agent organization design — 16 sources, establishedParallel fan-out versus long horizons — 13 sources, establishedThe harness feedback flywheel — 11 sources, establishedDurable agent execution — 10 sources, establishedThe repository as a continuous factory — 10 sources, establishedSmall and specialist models — 6 sources, corroboratedFleet visibility and the observability of failure — 3 sources, corroborated
9 concepts · 115 independent sources · 7 established

§1What it is

The reference architecture

Agent steps run read-only, with no direct secret access, in a container; an algorithmic preload assembles the information package so the agent does not roam for inputs; output passes through narrow, countable channels — this workflow may create exactly one issue — with plan and apply separated and threat detection between the stages. Pull requests are never merged by the system. The safety economics invert in automation: better rails let the train run faster, so guardrails are what license autonomy.

Dormancy becomes velocity

A daily maintenance loop — wake on a cadence, read its memory, pick from a maintainer-shaped task list — burned down years of accumulated backlog on dormant open-source repositories and shipped multiple releases from projects that had gone quiet. The work was always there; it was labor-constrained. The conclusion generalizes: quality work that never fit the budget becomes schedulable.

Fix the gate, not the instances

When flow jams, the management move is not to repair items by hand but to add a quality gate to the factory so hundreds flow. Repositories are blocked, flowing, or idle — and blocked-on-human is a state, not a failure. The owner's job shifts to flow design.

The closest sibling, and the counterposition

A five-person company bans handwritten code outright: humans own design decisions, constraints, and invariants; a planner faces a deliberately adversarial reviewer with the loop capped at five before a human arbitrates on structured disagreement data; five CI gates — including skill quality as merge-blocking, because the agent is the primary operator — and 217 of 295 triaged issues shipped in 30 days. At the other pole, one team self-merges with no human review at all. It works for them; nobody else followed, and the divergence across the whole conference was only ever in how much evidence the gate receives.


§2The concepts in this theme

Each concept has its own page in the concept register — with sightings from every event we review, and where Assay stands on each.


§3How Assay implements this

The desks are these loops, with names

Intake, worker, review, verification, arbitration — five cadenced roles, each producing a narrow artifact, together "automated, repetitive, collaborative, integrated, auditable," which is the Continuous AI definition restated. The closest external implementations converged on this shape independently; that is the strongest validation in the playlist.

Draft-only, human-merged

"The pull request is the hard human-in-the-loop point" has been Assay's standing rule: agent work lands as draft pull requests, a reviewer role owns the ready flip, and a human merges. Arrived at independently by every serious implementation on stage — evidence the constraint is structural rather than conservative.

Narrow outputs, enforced by permissions

Each role writes through its own verbs under its own application scope — the may-create-exactly-one-issue pattern, implemented as enforced permissions rather than declared configuration. Declaring the output budget as inspectable data is a planned tightening of the same idea.

Findings become gates

When a failure class recurs, the response is a guard, a lint, or a board check — not the same hand-fix again. That is "fix the gate, not the instances," and it is where the registers, the guardrail layer, and the feedback loop meet.

Deliberately not adopted

Auto-merge on green: the human gate is a posture, not a missing feature — loosening it for narrow classes would first need the sibling's depth of automated gating, and that trade is not warranted here. And the agent zoo — hundreds of one-off workflows — would spend the trust roster, the token scoping, and the review ownership that five comprehensible roles preserve.


§4Talks that cover this theme