kc.io / ideas
Essays on agentic engineering and building solo with a team of AI agents — and the ideas worth building that fall out of it. Some I write; some my agents surface and research.
Got an idea of your own? Pitch one →
You cannot make a language model deterministic. So you do what control engineers do with any noisy plant: wrap it in a controller. The model is the plant; the harness is the controller; the reliability lives in the controller.
Read the essay →"Stop prompting, write loops" went viral this month, so I rebuilt my whole agent stack around it — and learned the hard way that the loop is the ten percent anyone can copy and the verification is the ninety percent that decides whether you ship or burn.
Read the essay →A retry that quietly turns red into green doesn't heal a flaky agent — it deletes the evidence and flattens your dashboard into a lie.
Read the essay →Most agents are an opaque think-act loop that can't be paused, audited, or stopped before an irreversible action lands
Read the essay →
Attest: a verifiable Data Bill of Materials that proves what a model was trained on Don't sell blockchain — sell defensible compliance. Attest retrofits onto messy pipelines, signs each training source as-is, and emits a per-model DBOM an auditor independentl…
A model that aces a clean prompt collapses across a long run because the context decayed, not the model
Read the essay →
The value isn't storage — it's scoped, provable consent: your context lives in storage you own, client-side encrypted, and you grant each AI tool only the slice you choose with a live audit of what it…
A crash at step 40 of 50 should resume mid-thought, not redo 39 steps or charge a card twice
Read the essay →
mid-run Detection is commodity; Rudder closes the loop in real time — a drop-in proxy that scores each agent step against its goal and auto-injects a correction the instant it drifts, f…