I build the layer where domain knowledge becomes an environment an agent can actually work in.

• The environment — mounts a domain's operational context as structured state an agent navigates and acts inside. (domain-environment ↗)
• The judge — declares what a correct execution looks like and grades the actual against it, by decision not keystroke. (domain-verifier ↗)

The through-line: record what happened, declare what should be true, and treat the diff between them as the work.

Observe the real trajectories, separate the deterministic sub-decision from the judgment one, promote only the safe part to tooling, and leave the check running. Same loop at every scale.

Most recent, smallest scale: I mined 140 of my own agent sessions, found one wasteful pattern — unscoped recursive search — and routed it. ~250× faster, ~10× less context, zero model change, gated behind a 19-case test suite. The interesting part wasn't the speed; it was that 99.6% of the win was scope and consolidation, not a faster tool. Demonstrated in a narrow routing experiment: grep → git grep.

Head of Operations at Aroko. Writing and case studies at jeremycapps.com.