Insights

Field notes from the work — operator-level writing on the systems, decisions, and engineering trade-offs behind our ventures.

  1. We Re-Priced Leaving Our Forge. The Estimate Was Off by 100x.

    Every dependency carries an implicit exit price, and almost nobody re-runs the estimate. Ours was "a weekend plus months of missing features" — the real exit took an afternoon. Three switching costs collapsed at once, and the stale number was silently setting our tolerance for vendor pain.

  2. Interchangeable Programmers, Interchangeable Tokens

    Coding agents write better Go than Python and better Angular than React — and the reason isn't corpus size, it's corpus entropy. A narrow training prior plus fast in-loop verification is the sweet spot, and choosing a language for agent work is really choosing a corpus.

  3. A Function Named for a Lie

    strncpy was named for a safety contract it never honored, and it took the Linux kernel six years and 362 commits to delete it. Why the name was the original sin — and why deprecation-in-place fails when your code's readers are AI agents.

  4. The Half-Life of a Workaround

    Some of the scaffolding in a multi-agent system is structure the next model will still need; the rest is a workaround with an expiry date. A test — does the benefit survive more capability — and four habits for telling them apart.

  5. Reasoning-channel models eat your max_tokens budget

    Reasoning-class models bill their thinking against the same max_tokens ceiling as the answer, and code that sized for the answer alone now ships mid-sentence narratives to cache.

  6. Stop Paying for the Same Lesson Twice: Specialist Memory for AI Agents (Part 2)

    A month in, the ledger holds ~100 files across nine specialists — and three classes of lesson have already blocked repeat mistakes before human review.

  7. The Token Cost of Shorthand Is Real. Almost Nobody Puts It in the Right Place.

    ptal versus please take another look: the viral posts measure the wrong tokens, name a mechanism that doesn't exist, and still trip over a real effect.

    Also on LinkedIn

  8. Stop Re-Litigating the Same Lessons: Specialist Memory for AI Agents (Part 1)

    Two failed designs — coordinator-injected context and a monolithic runbook — and the per-specialist memory directory that finally pays for itself the second time you hit a wall. Part 1 of two.

    Also on LinkedIn

  9. The NULL Trap in Postgres Idempotency Locks

    A Postgres UNIQUE constraint with a nullable key column silently lets duplicate idempotency rows through, because NULL is never equal to NULL. The fix is an empty-string sentinel.

  10. Your Serverless Quota Check Has a Race Condition. Postgres Already Has the Fix.

    Two concurrent requests, one quota, and the bug that lets a paying user end up at 101 of 100. The fix is one line of Postgres — no Redis, no distributed locks, no retry storms.

    Also on LinkedIn

  11. From Chaotic Kittens to a Coordinated Squad: How We Cut Our AI Agent Squad in Half

    Six specialist agents became three, output quality went up, and compute costs went down. The lessons on context bloat, role overlap, and when to stop adding voices to the manager's head.

    Also on LinkedIn