Blog
The market sells workers. I write about the operating layer — gates, cases, and what the chrome actually shows.

Harness scoreboards are not a verdict
A morning spike in lines per token is not proof your agent harness got better. Cost, throughput, and quality have to move together — or you measured noise.

AI note taker, episode 1: I would not buy the ABH Pocket yet
ABH Pocket copies the Pocket spec sheet. The photos do not match. The open path is PKT01 plus pocket-libre.

The operating-system claim: rare, useful, easy to fake
An OS claim is a 15-of-220 shelf, and easy to fake. Memory inside it is ten gates, one question each.

OpenClaw when you already have Claude (or Claude Code)
Already on Claude Code? OpenClaw does not multiply your brain. Only a few pieces are unique: chat channels, always-on policy, and an optional local host.

Remote Claude Code: the laptop is not the runtime
Local Claude Code dies with the lid. Put the session on an always-on server; drive it with Remote Control from phone and browser.

Claude Opus 5.5: cheaper, faster, and four breaks
Opus 5.5 on Claude API: $4/$20, cache $0.20, four breaks — wiki to claude-api, claude-code, openclaw, litellm.
The map
21 drafts · not public articles yet
C1 — Category: why “agents” is the wrong noun (demand: scarce OS shelf)
- 001Why “AI agents” became wallpaper (and what replaces the noun)draft
- 005Boards are read-models, not the runtimedraft
- 006Observe → truth → propose → gate → humans decide (the loop as product spine)draft
C2 — Observe: nothing typed twice (demand: channels + docs; Stage-1 C: T03)
- 015Document intake is observation, not “DMS AI”draft
- 019Invisible demand: counting requests nobody loggeddraft
C3 — Truth: case, memory, graph (demand: brain/ontology 42; Stage-1 C: T04,T44,T38)
- 021Shared operational truth vs chat contextdraft
- 022The Case object: durable situation, not a threaddraft
- 028Approved baseline ≠ current timeline (the only honest slip)draft
C4 — Understand & predict (demand: anti-pilot 36; criteria 4–5)
- 034Lead-time exposure before site delaydraft
C5 — Gate & authority (demand: HITL 27 + govern 35; Stage-1 C: T13,T16,T34)
- 041Autonomy ↔ human authority: name both sidesdraft
- 042HITL as a runtime primitive, not a FAQdraft
- 046Stalled approval ≠ slow approvaldraft
C6 — Execute & orchestration (demand: orchestration 39; C02/C05)
- 051Closed execution loop vs recommendation dashboarddraft
C7 — Verify, learn, sleep (demand: eval stacks; Stage-1 enablers)
- 061Verify outcome ≠ HTTP 200draft
C8 — Interfaces people actually see (demand: market defaults to chat; our gap)
- 071Chat is the default product — and why that’s a trapdraft
- 074Front page as projection of cases, exposures, waiting decisionsdraft
C10 — Market language & GTM (demand: how to speak without mush)
- 091Five verb bets: Govern, Run, Work, Build, Knowdraft
- 098Measurement language over feature gridsdraft
- 100The publishing spine: how these 100 posts reinforce one thesisdraft
ai-note-taker
- —AI note taker, episode 2: PKT01 without the clouddraft
figma-prototype
- —Figma wins the board. Tokens win the paint. The library wins the chrome.draft