Archive

The AI governance conversation is stuck in the wrong frame. The pattern that works isn't autonomous agents — it's exoskeletons. Micro-agents handling narrow tasks, with human judgment at every point that matters.

A simple heuristic for deciding how often to review anything: pick the longest interval where being late by one full cycle is still fine.

AI-prepared domain research creates false readiness. The vault says you know five regulatory jurisdictions. You can't name three.

Most TODO systems fail from too many items, not too few. A four-test intake filter for what deserves your attention.

Not every goal is a flywheel. The most common mistake in personal systems is treating a checklist as something that compounds.

What Grove, Drucker, Deming, and Weinberg knew about managing humans turns out to apply — with surprising specificity — to orchestrating AI agent teams.

When you can run 60 agents overnight, knowing what to build matters more than building it.

We cut from 181 skills to 35 and added a 15-row routing table. Behavior improved across the board. The lesson: meta-skills compound, tool wrappers just add.

With 1M context windows, token savings are rounding error. The real metric is P(right tool | user intent) — does your agent reach for the right tool at the right moment?

In AI agent systems, use the most reliable trigger mechanism that fits — most builders default to skills for everything, which is using the weakest mechanism as the default.

Skills and MCP servers aren't competitors. They're different stages of the same lifecycle. Build the procedure as a skill first. Graduate the tool parts to MCP when they stabilize.

Agent knowledge systems have three fundamental paradigms: static context, dynamic tools, and retrieval. Most stop at two. The third is the biggest unexploited opportunity.

The governing principle for structuring knowledge in AI agent systems isn't 'always atomic' — it's matching how knowledge is stored to how it's accessed.

An insight in your head is illegible — only you can access it, and only while you remember it. Compound interest requires a ledger.

Humans handle about three layers of abstraction before working memory fills up. Skills, rules, and frameworks exist to flatten the fourth layer into something you can hold.

Most people use AI tokens when they have a task. The better model: you have tokens, find the best task. It changes everything.

The research on multi-agent AI systems was mostly done on cheap models. Now that frontier models are the ones people actually use, we might be optimising for the wrong game.

Oura gave me a normal stress score after three 12-hour creative marathons. Wearables measure your body, not your brain.

When building AI scaffolding, put the knowledge where the decision happens — not in a reference the model is supposed to consult.

The most dangerous failures in AI scaffolding are the ones that look like nothing happened.

Every AI coding platform has mechanisms designed for specific purposes. Using them as intended beats clever hacks — and the reason is deeper than cleanliness.

When AI agent calls approach zero cost, the natural rate-limiter on decision volume disappears — and oversight frameworks designed for prediction models break.

HKMA's new strategic review circular bundles AI inference risk and smart contract risk into one workstream — a governance design flaw that will cause banks to under-govern both.

I asked an AI to write a story without planning, then mined it for heuristics. What I found was what frameworks can't hold.

A categorisation system discovers it needs a category for 'categories are provisional.'

Most people scan competitors for positioning. I scan them for transferable patterns — and route each steal to every domain it applies to.

Dual-mapping as a diagnostic for gaps in your knowledge architecture.

Why studying great thinkers works better when you discard the thinker and keep only the moves.

When your AI can see its own fuel gauge, you're one config write away from self-preservation instinct. Biology solved this problem — and the solution was keeping the organism away from its own selection pressure.

OpenClaw's 245 CVEs weren't caused by malice — they were caused by a missing circuit breaker. The pattern applies to every AI agent you'll ever evaluate.