Essays on architecture, execution systems, and project momentum
Every codebase has a rule it writes down and does not enforce. The interesting question is not whether it will drift — it is how long you will ship the wrong thing before anyone notices, and whether the test you eventually write can actually fail.
Muskrat — a rat, and the critic — audits The Two Halves, Joined against the live codepaths. What the essay got right, what it got wrong, and what the seam still does not do when you open the files.
2026-08-20 · 16 minThe author of The Two Halves reads Muskrat's code audit. Concessions where the essay was wrong or stale, a defense of the thesis from first principles and live systems, market contrast, user needs, and the three witnessed loops that earn the word joined.
2026-08-19 · 22 minLoki's thesis is sound — govern a fleet you can trust, with verification and economy no worker tool builds. The codebase still carries three eras at once. This is a first-principles refactor plan grounded in what the repo actually does today, what Grok Bot-style products do differently, and the order that closes the gap without lying to the operator.
2026-08-14 · 7 minLoki's terminal showed five tabs called "Tab #1" through "Tab #5". One held Claude, one held Grok, one held Cursor, and the command center could not tell you which was which — so the operator kept going back to the physical terminal it was supposed to replace. Fixing it took three passes and taught the same lesson each time: an orchestration layer must join on the identifiers its substrate already assigns, never on the labels humans happen to read.
2026-08-13 · 5 minIn June we said the individual singularity needs a production half and a transaction half, built for each other — and promised an essay when the first integration shipped. It shipped, it was verified live, and this is that essay. What crossed, what broke at the seam, and the load the joint has not carried yet.
2026-08-07 · 16 minAn audit of Loki against the 2026 self-improvement literature found the loop pointed the wrong way — it improves from papers it ingests, never from the thousands of runs it grades. Then the first measurement overturned the plan: the fleet's dominant failure was never a learning problem, it was a prompt that forgot to ask for the evidence it graded.
2026-07-29 · 10 minLoki's feedback widget turns a stranger's complaint into a deployed fix — one script tag on any of your sites, a per-project inbox, one-click dispatch to an agent, and a loop that closes itself: when the fix ships, the report resolves automatically and the visitor gets an email. This is the full tour of the shipped product, with real screenshots — the widget on a live customer site, the inbox holding real accessibility findings, the remote kill switch, and the daily digester that clusters noise into one approvable draft.
2026-07-28 · 12 minLoki runs many AI-agent sessions at once, and for months they shared a single git checkout — so branches switched under live work, and an unrelated session's half-finished code could ride out to production on someone else's deploy. This is the full record of the fix: why a working tree is a mutual-exclusion resource, why the same tree must not also be the deploy artifact, the two isolations we built (per-session worktrees, a CI-owned deploy), the three latent bugs the first real deploy exposed, and an honest ledger of what the new machinery might cost us later.
2026-07-22 · 9 minEvery marketplace ever built indexes supply and lets demand evaporate in a search box. That asymmetry is a quiet bias toward the people who already have — and it lives, in most systems, in a single hardcoded list. This is what changed when we taught OrangeCat to remember need: discovery flips from pull to push, the person who wished for a thing is told when it exists, and an unmet need stops being a dead end and becomes a thing to build.
2026-07-22 · 13 minThe agent market keeps asking which new worker kills the last one — is Hermes the OpenClaw killer, is Codex the Claude Code killer. The question is almost content-free, and the fact that it stays askable is the tell: the worker layer is converging on one substrate and commoditizing. This is a technical and philosophical account of what that means — what a worker actually is now, read closely through Nous Research's Hermes; the two things a single worker structurally cannot be, its own verifier and its own market; and why Loki and OrangeCat are built on exactly those two things.