Zeptobeam

**This repo is an experiment.** We forked [ErlangRT](https://github.com/kvakvs/ErlangRT) and extended it toward AI agent orchestration — sync scheduler, reduction counting, supervision trees, ETS/DETS, hot code upgrades. It got to 343 tests and then hit a wall: the sync scheduler can't `await`, so every LLM call required two thread hops and correlation IDs. Reduction counting models the wrong resource — AI agents don't consume CPU, they consume tokens and dollars. The learnings shaped [ZeptoRT](https://github.com/qhkm/zeptort), which fixes both problems. Active development is there.

---

---

What

ZeptoBeam is the orchestration runtime for [ZeptoClaw](https://github.com/qhkm/zeptoclaw) — a multi-agent AI system where autonomous agents collaborate on complex tasks.

Built on Erlang/BEAM principles, it handles the hard parts of running many agents at once:

• Process isolation — Each agent has its own mailbox, state, and lifecycle
• Fault tolerance — Crashed agents restart automatically via supervision trees
• Message passing — Agents communicate via async messages, no shared state
• Preemptive scheduling — Reduction-based preemption with 3-tier priority queues
• Durable state — Checkpoints to SQLite, WAL-backed durable mailboxes
• Hot code upgrades — Swap agent behaviors at runtime without restart
• MCP integration — Expose runtime as MCP server, consume external MCP tools
• Advanced orchestration — DAG dependencies, retry policies, resource budgets, approval gates

---

The Problem

AI agents fail. They hallucinate