I write agent runtimes — the scheduler, supervision and state layer underneath the frameworks.
Most people building with agents work at the framework layer. I work at the layer below: process scheduling, supervision trees, state machines, and the verification gates that decide whether an agent is safe to turn on.
swarmrt — A native runtime and compiler for concurrent
programs. BEAM-shaped: lightweight processes, supervision trees, message passing, preemptive
scheduling. 55K lines of C across 96 files, zero dependencies. Plus sw, the language that runs
on it.
swarm-code — A terminal coding agent written in
sw, running on swarmrt. No Node, no Python.
carbine (private) — A BEAM-native coding agent that runs on your own endpoint. Bring-your-own-model, whitelabel-ready.
o — An autonomous software loop. Watches its own board, dispatches coding agents, verifies with real tests, auto-merges green work, auto-reverts regressions, budgets itself daily.
swift-render — Programmatic motion graphics in Swift. SwiftUI scenes + Metal shaders → MP4.
- ex-Head of Edge AI, Sarvam AI
- ex-CTO, Collectiv AI — shipped a coding agent in 2023, on GPT-3.5, back when the model would not reliably hold a tool schema, keep call ordering straight, or preserve state across turns. The three failure classes below are ones I fixed by hand back then, before any framework existed to hide them.
- Building in AI since the GPT-2 evals days
- Now at Otonomy Corp — autonomy stacks for GPS-denied navigation and onboard perception
If your agent works in a pilot and stalls in production, it is almost never the model. It is one of three things:
- Context assembly — the fact it needed was absent, buried, or evicted before the model saw it
- Tool arguments and ordering — right tool, loosely-typed argument, or an ordering constraint that lives in the prompt instead of the schema and fails under load
- State across turns — two tools mutating the same object, half-applied on retry, so the model is handed state that no single tool wrote
Send me one failing trace and I will spend 90 minutes telling you which one it is. Free, no deck. If the finding is worth acting on we can talk about the fix; if not, you keep the finding.