Maturity, not complexity. Mature AI-first setups look intimidating from the outside. Conceptually they're the same as anyone's – folders, markdown files, an AI that reads them. The dimension that matters is maturity: of your context, of your workflows. When you start, there are inaccuracies, contradictions and gaps everywhere. Catching and fixing them over time is the discipline, and it's hard to short-circuit. The answer is just to start. I recorded a 30-minute version of this for a founder community – blog write-up and a clip series coming.

Notion is for you. The vault is for agents. Two readers, two formats. Notion is human-edited and visually structured. The vault is dense markdown, AI-managed end-to-end. People running both in parallel keep finding it awkward. The fix is deciding which file is for which reader, then writing for that reader only.

CLAUDE.md is the map. The docs are the territory. CLAUDE.md (persistent context file an agent reads at the start of every session) is still context – just the orientation layer, pointing the agent at what to read next. The dense detail lives in the files it links to. Working backwards from rich detail is much easier for an agent than filling gaps from a sparse start. Mine barely changes. The docs it points at change constantly.

Stop pushing to 1M context. Close sessions at 200–350k tokens. Push to the limit and you hit caps faster and performance degrades inside the same session. A clean exit beats a long thread every time. A solid CLAUDE.md + memory bank setup makes the next session free. Blog post

Cowork is a personal assistant. Claude Code is how your business runs. The "is Claude Code just the technical version of Cowork?" question came up several times this week. Cowork has skills and project instructions, which is what makes the comparison feel close. But it's missing the four things that make a setup compound: custom sub-agents (delegation), hooks (guardrails), replaceable system prompts (sharable team personas) and parallel sessions (conductor mode). Each one blocks a different layer. Blog post

The gap is widening. I spoke at a company offsite Mon–Wed. The pattern in the room was the same one I keep seeing – real appetite and excitement around what AI can do, matched with frustration about not knowing where to start. Meanwhile the gap between people using this and people not is getting bigger by the week. The fix is just to start. Maturity comes from iterating on something live, not from getting the structure right on day one.

On my radar

MCP vs CLI was the technical thread of the week. Most 'MCP is bad' takes are arguments against badly designed MCP servers, not against the protocol. CLI wins for ad-hoc work and same-session evolution. MCP (Model Context Protocol, the standard agents use to talk to external tools) wins where governance, telemetry and multi-agent coordination matter. Both have their place. AI Engineer EU talk

The Max plan squeeze is real. Anthropic announced claude -p (the non-interactive flag used for production agent pipelines) is being removed from Max plan accounts. Combined with reported ~35% token bloat on Opus 4.7's tokeniser, the gap between subscription and API economics is closing fast. If you've built agent infrastructure on Max, plan the migration now. Announcement