Environment priming and substrate readiness

An agent that spawns instantly but cannot do useful work for forty minutes has not started. The metric is time from dispatch to first productive action, and the fix is pre-primed cached clones rather than cold builds. The second layer is where failure is reported: failing tools, tests, lints and builds are the highest-leverage place to deliver context, so error output must carry a stable machine-routable code alongside remediation text — why it failed, what to do instead, where the runbook is — letting the agent self-correct without a human.

Human-shaped surfaces force agents to reconstruct sequence and parameters from fine-grained endpoints, then poll, guess and retry. Agent-native surfaces expose one-purpose, non-overlapping, state-aware capabilities written for a model reader; they emit events instead of demanding polling; and at the UI boundary they replace screenshot-and-guess with a declared tool contract. The design rules that recur: one purpose per tool and never overlapping, register only what is valid for the current state, plain-language descriptions rather than marketing copy, strict errors carrying technical detail so the agent can iterate, small outputs returning exactly what was asked.

The repository layer is where the effect is now measured rather than asserted. Repositories with no agent configuration get worse after adoption — more cognitive complexity, more static warnings, more duplication. Basic configuration reverses that: PR throughput up with revert rate, duplication and complexity all down. Structured context is where codebases degrade several times less at the same or better speed. And the recurring corollary is worth as much as the headline: designing for agents turns out to improve the human surface too.

Also: the Stanford readiness-level classifier (open source) and the IEEE ASE paper due in October; WebMCP as a W3C proposal, Chrome DevTools MCP and Puppeteer; the AIDaR data-readiness framework and the NeurIPS Paris workshop on data infrastructure and evals for agentic AI.

Spec plus skills plus guardrails being the difference between agents helping and harming a codebase is the core claim of the public bundle, and it now has an outside measurement behind it rather than an argument. The concrete action is external calibration: run the lab's open-source classifier against Assay's own repositories and see which level they actually score, rather than assuming the top one.

The same forty-minute tax appears whenever a worker claims a brief and then spends ten minutes resolving dependencies or regenerating the board. Worktree management and session boot cut it, but nothing measures claim-to-first-productive-tool-call — which is exactly the metric this evidence argues for. Error contracts are further behind: desk verbs and guardrail hooks emit human sentences today, where the frontier position is that every refusal should also emit a stable code an agent can route on — retryable versus STOP versus escalate. That matters here because a guard refusal is a hard stop, and an agent that cannot tell a stop from a retry will route around it. The falsifiable prediction is fewer follow-up agent turns after a mid-loop verb failure.

Desk verbs are already intent-level operations rather than endpoints, so that shift landed independently. The gap is at the edges — anywhere a skill body tells an agent to hand-roll a raw forge call is an endpoint-shaped leak, and the five model-consumer rules are worth auditing every verb's help and error text against. One piece is deliberately rejected: no event bus. Dispatch claims poll git on purpose, and no observed failure needs a push channel. The UI-contract line stays an unbuilt experiment — expose two or three read-only diagnostic tools on one high-value internal page and gather evidence through a declared contract instead of the screenshot path, measuring tokens and wall-clock on both — with the hard caveat that the browser API behind it is single-vendor, flag-gated and non-headless, so it can never become a load-bearing CI rung. "From inference to a contract" is also a citable framing for why briefs carry explicit definition-of-done contracts.

• The IEEE ASE paper when it lands in October — whether the readiness-level effect sizes survive peer review, and whether the open classifier can be run against a private repository.
• Whether machine-routable error codes appear as a stated convention in an agent harness or CLI toolkit, rather than one vendor's log redesign.
• Whether WebMCP progresses past origin trial and gains headless support; without it the UI-contract path stays a developer-session convenience.
• Whether anyone publishes a time-to-first-productive-action number, which would give the priming claim a metric instead of an anecdote.