Multi-agent orchestration harness
2026
Independent project · solo
- An orchestrator delegates real work to worker agents — long-lived, file-backed sessions in isolated child processes, with concurrency caps, per-agent error isolation, and resume across restarts
- The part I care most about: the rails. Every worker sandboxed in its own git worktree behind a runtime path-jail; a shell runner that passes args as an array, so a branch named
feat/foo; rm -rf /is data and not a command; an allow-list gate kept in lockstep with what the agents are told - No agent approves its own work — the PR lifecycle is refused to workers and routed through a path that owns push, review, and merge. Closing that door meant closing the side doors too: mutating API calls and label edits, since labels gate merges
- A local CI system issues per-commit verdicts — a red run automatically becomes a diagnose-only investigator agent, able to read and reason but not to fix
- Zero-dependency Node/ESM services plus a TypeScript extension layer, bridged to the browser over JSONL and Server-Sent Events
- Also inside: on-device speech-to-text, token-usage accounting with provider-limit backoff, event-sourced ledgers whose analytics are derived rather than hand-kept, a wikilinked knowledge graph