Agent harness design, context engineering, agentic infrastructure, and the emerging shape of agentic organizations. Practitioner-positioned, research-grounded.
Agents are constitutionally bound to their principals. When they transact on your behalf, they optimize for your interest — not Visa's, not Stripe's. That makes the 30% intermediary fee not just expensive, but architecturally wrong.
The agent harness determines the behavior envelope, not the model. Six agent harness architectures — and why the sixth breaks what the others assume.
The AI productivity discourse optimizes for the prompt. The actual unit of knowledge work is the thread -- and nobody is building for it.
Anthropic named a memory consolidation feature 'Auto Dream.' The metaphor wasn't decoration — it was a spec. Wittgenstein explains why.
Anthropic's Claude Code source leaked — 512K lines, every agent pattern exposed. The model layer fell to distillation. The harness layer fell next.
Permissionless Intelligence — The Agent-Native Currency
2026-04-17•8 min read•analysis
The Thread Is the Unit — Knowledge Work After the Prompt
2026-04-14•9 min read•essay
Agent Harness Taxonomy — Six Architectures for AI Agents
2026-04-07•12 min read•analysis
The Trust Contraction — When Trust Is the Only Scarce Input
2026-04-07•7 min read•essay