Meta's ScaleMAS Uses a Searchable Message Store to Fix Multi-Agent Coordination Overhead
RobertTLange · x · 2026-10-24
ScaleMAS (Meta FAIR + Harvard, Zhenting Qi et al.) tackles why adding agents to LLM multi-agent systems adds coordination overhead rather than pure gains: growing reading burden, useful work buried, fragility to failed agents.
- Agents interact with a shared datastore of persistent, linked messages via three ops: publish (with optional references), content search, and visit-by-address
- Search returns at most k results ranked by relevance and citation count—citations proved an effective quality proxy; messages outlive their producers
- Across four tasks (BrowseComp-Plus, FATE-M, SWE-Bench Verified, Contradiction Audit), ScaleMAS beats agent-centric and blackboard-style baselines at lower inference compute, scales to 128 agents, and tolerates mid-run agent failures
More from coding & agent
- Intent Mobile launches: steer hundreds of coding agents from your iPhone — LukeW · 2026-10-24
- GPT SOL 6.1 builds a 3D axe from scratch via Blender MCP in 7 minutes — OpenAIDevs · 2026-10-24
- icons.computer indexes 3,000+ curated app icons, agent-ready via Firecrawl — devdigest · 2026-10-24
- Dev: With LLMs Helping Setup, There's Never Been an Easier Time to Daily Drive Linux — KevinNaughtonJr · 2026-10-24
- Postman's first AI-Ready API leaderboard puts You.com (96) above OpenRouter (94) — RichardSocher · 2026-10-24
- Richard Socher launches Flint, an open-weights local coding model hitting 125+ tokens/s on a MacBook — RichardSocher · 2026-10-24