Shared memory · loops · evals — for coding agents
One shared memory for every agent on the team. Scheduled loops for the recurring work. Evals that prove — not promise — what actually helped.
Loops run and leave records. Evals score the runs. Learnings that held up under scoring get promoted into shared memory — and the next runs start sharper. Each pass removes a little roughness; the compounding is the product.
The stone turns; the stations hold still. Loops feed run, evals own score, memory owns learn — and promotion is the eye of the stone: only learnings that earn their score become canon.
Loops · run
Nightly audits, PR reviews, weekly digests — recurring agent runs become first-class records: what ran, what it cost, how it ended. The mill turns while you sleep.
Evals · score
Deterministic asserts and judge rubrics, runnable locally and as a CI gate. Every score is pinned to a commit and a model, so improvement is a fact — not a feeling.
Memory · learn → promote
Every agent on the team reads and writes the same memory. New learnings land in an inbox with full provenance; only reviewed, evidenced knowledge becomes canon. No drift, no poison, no “my agent knew that.”