Blog
Short write-ups from actually running this stuff — what held up, what broke, and what we'd change. A summary here, the full write-up one click away.
Classify by the interaction edge, not the symptom, and trace back to the earliest failure nobody recovered from. Plus the six checks worth running before you blame the model — and where the new taxonomy is weaker than it looks.
A status report generated by the same process being checked isn't evidence — whether that process is a code review or a health check. What our own fleet's audit caught in nine days, and what slipped through when the check was missing.
An orchestrator started mixing up its own tasks mid-run. The instinct is the kill switch — and it's the expensive move. What atomicity, sagas and SRE load-shedding say about stopping a fleet safely, and the cases where you hit the switch anyway.
A factorial study, a presence-effect RCT, and our own fleet's task history, argued through to a specific design: what actually earns a place in an agent's instruction file, and the two-variant template built from it.
Sessions can now run in their own git worktree instead of a shared checkout. What a worktree actually buys an agent, the real disk/complexity cost, and the internal incident that decided the cleanup mechanism had to ship in the same release as the isolation, not later.
Running one Claude Code or Codex session is a solved problem. Somewhere around the second or third at once, something shifts — and it's usually not the agents themselves that start costing you time.
A live benchmark of agent-memory systems against a bare-model floor and a flat-file baseline — 6 live-tested, 7 tiered. One leader (narrowly), one real surprise, no vendor claims taken at face value.
Our standing rule for the harness equipment registry: a claim from a summary or a README never gets cataloged as fact until it's re-confirmed against the primary source. What that looks like in practice, and why we hold our own product to the same bar.