Writing

keeping the caddy's evals honest

the rules around bagman's eval suite: lint before a model call, treat the ci workflow as a contract, keep safety cheap and deterministic, fix the fixture instead of the threshold, and pin every bug the evals find.

4 min read · → bagman

custom architecture checks as guardrails for coding agents

coding agents copy whatever pattern is nearest. i wrote 67 architecture rules into my cli so the nearest pattern stays correct, and a verification gate that fails when it checked nothing.

8 min read · → swarm

the evals behind bagman's caddy

how bagman's caddy is tested against real models: one scenario per behaviour, eight scorers, suites that gate and suites that only measure, a weekly model sweep, and the regression check that keeps a context bug from coming back.

7 min read · → bagman

making the site readable to agents

scanners like is-agentic and ora now score whether an agent can find, fetch, and use a site. this is the surface bokendell.com actually ships for that, without inventing a rank we have not earned.

4 min read · → portfolio

how the caddy's prompt is layered

bagman's caddy picks a model per role, then builds its prompt in four layers so the parts that don't change during a round are cached for an hour. why the layers split where they do, and the tests that keep the cache from breaking.

7 min read · → bagman

⌘Kterminal⌘Lask