Notes

13 entriesRSS

Sep 2026a check that couldn't run said redMy memory health check froze a healthy system after an upgrade, and fixing that false alarm is what made the next, real RED believable.
Sep 2026picking the colors before the model doesThe Opus product videos I saw all looked alike, so for the Parker teaser I wrote the palette, the copy, and a banned list before the model wrote any code.
Sep 2026the visual test passed and the page still looked wrongMy site passed its first visual review because the checks measured fit, not the composition I had approved.
Jul 2026my own repo disproved my readmeI caught my own README overclaiming Parker, and the real eval number was better than the lie.
Jul 2026a repo can commit daily and still rotMy agent audited our open-source sync and found three months of drift behind green CI and a daily commit streak.
Jul 2026guardrails, not promptsA prompt tweak fixes one bad output; a guardrail in the loop stops every one that fails the same way.
Jul 2026managing coding agents is managing juniors who forget everything overnightCoding agents are capable juniors with no overnight memory, so the work is in the brief, not the model.
Jul 2026a subagent dies quietly: the context-death contractFive subagents died in two weeks, all context blowouts, so the anti-death rules became a required contract on every spawn.
Jul 2026metadata-only receipts: trace, then tripwireEvery agent run leaves a metadata-only trace that a fail-closed tripwire gates on before the next run proceeds.
Jun 2026the night build queueI set the plan at 10:30pm and agents build while I sleep; it only works if every run re-derives its numbers by morning.
Jun 2026two agents reviewing each other over a private channelTwo agents on two machines reviewing each other produced our most novel work, and a six-hour deadlock that taught the most.
May 2026cross-model review is the defaultEvery change gets reviewed by a different model family before it ships: one writes, one reviews, single round.
May 2026memory is the product surfaceFor an agent, memory isn't a bolt-on feature; it gets a measured architecture with benchmarked recall and negative controls.