Writing about AI, infrastructure, and what actually ships.
total 141 · ~10.9h read · updated Jul 02
Principles 21–25. The governance and safety layer: strictKnownMarketplaces, no goal-conflict prompts, quarterly AppSec, four telemetry signals, monthly incident discipline.
Principles 15–20. The harness configuration that keeps the kernel and lifecycle cheap: CLAUDE.md under 200 lines, hooks for real incidents, skills that auto-invoke, subagent isolation, pinning, and Stage 5 distribution.
Principles 6–14. How work moves through an agentic engineering team: the ticket as contract, AI distillation with human curation, three gates, verification before done, characterisation tests, the 1.2× capacity rule, the J-curve, and telemetry.
Principles 1–5. The five rules that everything else in the framework rests on: standardise the harness, make verification load-bearing, default to plan mode, pick the cheapest layer, reflect every task.
Practitioner consensus puts the cliff around fifteen tool calls per prompt. Here's why agents degrade past that, and the three operational rules that keep them on the safe side.
Anthropic's multi-agent Research feature beat single-agent Opus 4 by 90.2% — at 15× the token cost. Every documented production swarm runs on rails. Here's the topology decision framework before you commit.
Agents over-refactor stable code without a safety net. Feathers' characterisation-test technique — write tests for current behaviour before changing anything — is more important than ever. The agent itself is the perfect characterisation-test-writer.
Karpathy named one mode. Willison named the other. Most 'AI failed in production' stories are actually 'we promoted a vibe-coded prototype without transitioning into the production discipline.'
METR ran the experiment. AI made experienced developers 19% slower — and they reported feeling 20% faster. The week-6 dip is the bottom of a documented J-curve. Most pilots get cut here. The right ones don't.
AI is making junior output look senior-level while preventing junior skill from forming — and the Stack Overflow collapse just removed the ambient learning layer that used to catch the deficit. Three interventions that work.
ReAct gave us a three-step loop. Production hardened it into five. The two new steps — Plan and Verify — are where everything that goes wrong, goes wrong. And the field has now named the worst offender.
Three open-source extractions converged on the same five layers. The architecture isn't a vendor narrative — it's a discovered structure. Here's the decision rule that keeps you from over-engineering it.
