Measure outcomes, not activity

Merge rate — around 74% across the studied population, and explicitly the target, because 100% means rubber-stamping. Zero-touch merge rate: 84% of merged agent PRs landed without a single human fix commit, the honest measure of "good, or just loud?" And turns-to-correct: how many rounds an agent needs to get it right, which measures the harness rather than the model.

Of every hundred dollars spent on AI tooling, roughly eighteen becomes code shipped to users; the rest is rework and generated bugs. Organizations that simply bought licenses for everyone are retrenching. The failure pattern is consistent: generation scaled, review and verification did not, and the queue moved nowhere.

With no skill installed, agents still follow the skill's instructions 55% of the time — the knowledge is in the weights. Skills earn their context tokens almost only on house-specific knowledge: approved libraries, required sequencing, prohibited patterns. The measurement discipline: activation (did it fire?), trajectory (was the workflow followed?), outcome (did it work?) — three separate numbers, most teams measure only the last.

The longitudinal finding underneath everything: repositories with no agent configuration get worse under agent use — complexity, warnings, duplication all rise. Repositories with structured context degrade roughly three times less at the same or greater speed. The tooling is not optional overhead; it is the variable that decides the sign of the outcome.

Eighty auto-opened pull requests become noise nobody reviews. The working pattern scores findings by return on effort, automates the investigation that precedes a fix, and keeps the human gate. And agent workloads are measured as distributions with defended percentiles — never a single-run average.

The status board is derived from authored artifacts with consistency linting and independent re-verification — and the specification forbids claiming more than that. Honesty about what is measured is the precondition for measuring the right thing: a dashboard that overclaims cannot be corrected by its own numbers.

The lifecycle separates implemented from verified from done, and requires evidence for each transition. Rejected verification, review rounds, and gate dwell are states on the board — the raw material of turns-to-correct, already recorded, not yet computed.

Assay's operational reporting still leans toward activity. The outcome numbers this concept calls for — merge rate per role, zero-touch merge rate, turns-to-correct per brief, activation per skill — are not yet produced, and all four are computable from data the methodology already collects (agent identities make zero-touch merges trivially detectable). This is the active roadmap item, stated as such.

The configuration-level finding — unconfigured repositories degrade, structured-context repositories hold — is the empirical case for what Assay ships. It converts "process overhead" into the measured variable that decides whether an agent fleet helps or harms a codebase.

8 of 39 talks shown — the ones that reach the most concepts in this theme. Every sighting, per talk, is on the concept pages above.