One warm context beats five cold ones
Per-layer workers each pay full context assembly. Loading once and generating inline amortises it — and the measured saving was smaller than estimated.
codegen
12 posts
Is this derived thing current? Make it a comparison
Staleness checked by timestamp or by memory fails quietly. A digest of the declared inputs, stamped beside the output, makes currency a comparison.
Which artifact is actually the design
When several artifacts describe one screen, only the highest-fidelity one is read by codegen. Editing a lower tier looks like a change and reaches nothing.
The running app and the spec disagree, and only one is documented
Hand-editing generated source creates a divergence from the spec that nothing reports and the next regeneration silently erases. The fix belongs upstream.
The orchestrator that wrote the code itself
When the coordinator writes a layer instead of dispatching its generator, the conventions the generator carries are gone and the output still compiles.
Editing the generated thing is how the spec stops meaning anything
A change implemented directly in source skips the spec, so every report downstream is computed without it. The feature works and nothing can see it.
A generator given the wrong context does not stop
Scoping a context pack by project type saved 8.2% per subagent, not the 40% I guessed. The real risk is a thin pack: a missing surface makes a generator invent.
The corpus your generator reads is production code
The corpus a generator reads is as load-bearing as the code it emits. A hand-edit to it has no provenance and is dropped by the next re-sync.
An agent-driven refactor, with the receipts
Before and after on a real module migration: what proved the change was inert, what the agent got wrong, and the receipt that made it safe to run at all.
The same rule in two files is how agent-written code drifts
A protected-path list was typed by hand and had drifted nine entries behind reality. Derive the list instead of copying it, and gate the copies you keep.
Three kinds of code I stopped letting an agent write
An honest negative result: three categories where generated code cost more to verify than to author, and the property they share that predicts it.
100 technical posts, and the gates that stopped them being slop
Publishing at volume with an AI agent is easy. Publishing at volume without shipping slop needs machinery that refuses, and refuses loudly.