Humans move up-stack

Agent access widens the spread between people rather than compressing it. The division that held: humans own systems thinking and design judgment; agents excel at policing drift against a human-owned record. Design documents written by agents failed everywhere they were tried — until humans authored them and agents enforced conformance.

The largest case study — five hundred people, weekly usage above eighty percent — ran on a maturity model delivered as recurring self-assessment workshops, where the discussion mattered more than the score, and on small repeatable patterns spreading by word of mouth. The most consistent question across every workshop: can we trust the code? The org-math floors held elsewhere: on-call rotation arithmetic puts a hard minimum under team size, and unread generated code pushes review left — to the specification.

Open source's human-to-human contract broke under parallelizable free code — one prompt produced ten nonsense pull requests in under a minute. The controls that worked: requiring an issue first (it filters roughly ninety percent of drive-by contributions, because agents do not come back to file one), rate limits on external contributors, checked-in context files for every agent, and vouch/denounce ledgers — a reviewed file of who is trusted, enforced in CI, with field data showing quality up and volume retained. Agent identity is the load-bearing primitive underneath: an agent with its own identity is an access-control boundary, and attribution is what makes review honest.

Cognitive surrender: waving through the too-big change, defaulting to the largest model, skipping the sandbox, "I can't work on this codebase without AI anymore." The management question the track left behind — if you lead, are you building an environment that leads to surrender?

Humans author briefs and rule on forks; decision requests are closed by humans only, with the ruling recorded on the record. Agents execute, and police drift against what the human wrote. Judgment lives in prose; claims live in code — and the two are kept separate on purpose.

Nothing in Assay merges itself. The gate is the human's control point, evidence-bearing rather than ceremonial — which is the amplifier's volume knob, and the answer to "can we trust the code?" that the adoption workshops kept arriving at.

Every action lands under a role's own identity; the registers carry who acted and why; merges preserve two-parent history an auditor can walk. The conference's trust layer — identity, attribution, ledger — is the part of this concept Assay has shipped rather than discussed.

Worth writing down: the autonomy ladder — Assay runs graduated autonomy in practice, from human-driven loops to one narrow verifier carve-out, but the ladder itself is unwritten, and adopters would benefit from seeing it. Mistake replay as regression tests, and risk-and-size triage labels at the gate, are cheap imports from this track. Not adopted: the community-building apparatus — Assay's adoption problem is outward-facing, and its "community" is the bundle plus the trust roster.

8 of 45 talks shown — the ones that reach the most concepts in this theme. Every sighting, per talk, is on the concept pages above.