The Org Reshapes, Not Shrinks
Katelyn Lesse (Head of Platform Eng, Anthropic): same headcount, several times the parallel projects, more PMs + Prod Ops, no dedicated QA, define the outcome, and a brain/hands architecture. AI-Native Eng Org, Part 3.
Personal Knowledge Base · 2026
From Agent architecture to enterprise strategy, from Agentic Coding to safety boundaries. A Solutions Architect's AI field notes.
How resource constraints forge human judgment
An original cognitive-science piece. The human brain compresses experience far harder than LLMs, and Taste lives in that gap. Four research threads — Gigerenzer, Webb, Friston, resource rationality — assembled into one causal chain: constraint → tradeoff → structural representation → directional intuition, with explicit boundaries and falsifiability conditions.
Read the essay →The complete manual for the ready-made framework
A consolidation of a dozen-plus field notes. The model is the fulcrum; the harness is the lever — a five-layer build order from picking a model and managing context to knowledge tools, orchestration, and verification, plus onboarding a 700K-line legacy codebase. Configuring it, not building from scratch.
Read the article →The decisions a framework makes for you, you make yourself
The first piece covers configuring ready-made Claude Code; this one covers building that harness layer from scratch on the API/SDK. A dozen design-side reports gathered into three blocks — tool design, orchestration, verification & security — plus one prompt-caching constraint running through all of it. Almost zero overlap with the first: consumer side there, design side here.
Read the article →Why "playing it safe" is slow-motion elimination
At the opening of the AI shift, most leaders still treat a bet they should go all-in on with a lottery-ticket mindset. Anchored to McKinsey's power curve, the autopsies of Kodak/Nokia/BlackBerry, and $416B of real capex — one ruler, "does it hurt," to tell a real bet from a show.
Read the essay →Based on Anthropic's ICLR 2026 paper. AI organizations are more effective but less aligned than individuals. Three failure mechanisms.
Three integration patterns, five design patterns, Client cost reduction -85% tool tokens.
Multi-dimensional analysis and tiered framework for Agent autonomy.
Design-level thinking on harnessing Claude's intelligence.
MacCoss Lab 700K C# codebase: standalone context repo, Skills reference-not-embed.
8 major trends: the evolution from Copilot to Autonomous Agent.
Context rot, compaction, rewind, subagent decision framework.
Scheduling, collaboration patterns, and isolation strategies explained.
Skill authoring guidelines, pattern library, and reuse strategies.
Cache as an architectural constraint, not optimization. 5 counter-intuitive designs + strategic analysis.
Managed vs direct: cost, latency, and feature coverage comparison.
API / Claude.ai / Claude Code privacy model differences.
MRCR v2 collapse (256k 91.9%→59.2%), BrowseComp −4.4pp.
Tokenizer changes, xhigh effort, adaptive thinking, 3 behavioral changes.
L'Oreal / Lyft / RBC trust-first approach and measurement framework.
Enterprise AI security defense strategies and implementation paths.
AI safety threats, training monitoring, sandbagging — 15-page deep analysis.
Solow Paradox 2.0: $250B investment vs 10% output. 30+ data sources.
How resource constraints produce Taste; Less-is-More effect; human-AI cognitive stack collaboration.
Classic six factors + GenAI's seven paradigm shifts + weighted scorecard.
Five signals: Vertex to Gemini Agent Platform, 8th-gen TPU.
Key insights and trend distillation from 19 sessions.
Design methodology showcase and practice.