@metaharness/radio

Passive-awareness swarm bus for agent pods. Freeze the model. Evolve the comms policy.

A dependency-free, deterministic TypeScript implementation of AgentRadio (arXiv:2607.28430) — the asynchronous messaging layer that lets multi-agent systems communicate during task execution instead of only at phase boundaries. See ADR-241.

Why

The paper's ablation on long-horizon codebase QnA (four agents vs one, same model): naive division of labor +7.2 pts, negotiated partition +12.1, and passive awareness +10.5 (p = 0.0023) — 62.1% vs 32.3% single-agent overall. The passive layer is one primitive: wait_for_mention backgrounded as an OS task, not an LLM step. Listening becomes free; an @-mention folds in at the next step boundary with a full thread snapshot, never interrupting a running command.

What ships

• RadioBus — threads, non-blocking send with @-mentions, total order by a logical clock (no wall clock: replays bit-for-bit).
• Watcher — fold() surfaces pending mentions + snapshots at a step boundary at zero step cost; blockingReceive() is the ablation arm and costs one. Separate methods, so a protocol can't silently mix the modes.
• runProtocol — the paper's five-phase, assembler-gated pod protocol: Explore → Divide (negotiate until approved) → Execute (live sharing under passive awareness; silent under blocking) → Review (can reopen Execute) → Submit (unanimous).
• runSim / makeTask — a seeded, deterministic codebase-QnA swarm sim (facts scattered across units, configurable cross-partition fraction). Its headline property: reproducing the paper's ablation order passive < negotiate ≤ divide < single in foreground steps-to-resolve.
• scripts/flywheel-radio.mjs — evolves the comms policy {mode, foldEvery, postPolicy} from a deliberately bad root under @metaharness/flywheel's frozen gate + never-optimized anchor topology, Ed25519-signed and replay-verified. The wheel re-discovering passive/immediate is the paper's ablation direction, measured on our own machinery.

Honest bounds

Mentions can derail (the paper's passive layer gained 47 rubrics, lost 23); awareness cannot surface conclusions no agent forms; the sim is a mechanism testbed, not a benchmark claim. Live LLM-pod wiring is deferred to its own measured ADR.

Usage

import { RadioBus, Watcher, runSim, makeTask } from '@metaharness/radio';

const bus = new RadioBus();
bus.createThread('worklog', ['scout', 'builder']);
const scout = new Watcher(bus, 'scout');
bus.send('worklog', 'builder', 'auth flow lives in gateway/, not api/', ['scout']);
// ...later, at scout's next step boundary — zero step cost:
for (const { mention, snapshot } of scout.fold()) {
  /* fold the discovery into the current sub-task */
}

MIT.