Agent infrastructure

An agent’s output is bounded by the environment it lands in, so runtime, feedback channel, interface and data layer are assessed and invested in up front rather than diagnosed after a failed eval. The measured result: spec plus skills plus guardrails is the difference between agents helping and harming a codebase — readiness levels, externally classifiable.

Multi-turn tool-calling agents present an ever-growing shared prefix, very long idle sessions and unbounded token demand — the serving stack reorganizes around the KV cache. Interop protocols converge on statelessness and minimal envelopes with governed context handoff. And retrieval built for human search is the wrong shape for agent traffic: a different query distribution, specialist domains generic embeddings cannot separate, and no first-class doors for agents.

Computer-use and web agents drive the human interface as the integration surface of last resort — a model perceives the screen and acts through the channel a person does, adapting when the UI changes. The distinction from RPA is adaptation.

Desk verbs are already intent-level operations rather than endpoints, and every refusal a guard emits carries a human sentence. The readiness-levels result is the strongest external corroboration yet of the core claim — spec, skills and guardrails being what makes agents safe to point at a repo — and the concrete action it prompts is external calibration: run the open-source classifier against our own repositories rather than assuming the top level.

The minimal-envelope bet matches: the desks interoperate through markdown, git refs and forge primitives, with question, help wanted and needs-decision as the clarify-instead-of-fail vocabulary — the same position the working group reached after abandoning stateful in-band elicitation. Behind: error contracts (every refusal should also emit a stable, machine-routable code — retryable versus STOP versus escalate), and claim-to-first-productive-tool-call is unmeasured. No event bus, on purpose — dispatch claims poll git and no observed failure needs a push channel.

Serving infrastructure sits below the API line and is deliberately out of scope, though its second-order consequences — longer desk sessions stressing claims, worktrees and crash recovery — are noted. Retrieval is path, filename and grep today; the pre-commitment the evidence buys is to build a discrimination benchmark from our own ontology before trusting any similarity score, because a confident citation of the wrong finding is worse than no citation. Computer-use does not bite where real APIs exist.

8 of 29 talks shown — the ones that reach the most concepts in this theme. Every sighting, per talk, is on the concept pages above.