Swarm sweeps: N agents + shared forum + arbiter for coverage tasks #426

From Anthropic's 'Patterns and problems in emerging multi-agent systems' (Aug 2026): 45 agents, each a VM, a SHARED FORUM to coordinate, and an arbiter agent deciding validity — 266 vulnerabilities found vs 21 for uncoordinated parallel agents (complementary coverage; the swarm specialized and built its own tools).

For Solenta: a swarm automation pattern for coverage tasks (vuln scans, dep audits, large refactors): N worker threads on one task, a shared forum artifact they read/write (the memory server is the natural substrate), and an arbiter thread that dedupes and validates findings before they surface. Distinct from best-of-N (same task, race) and workers (one orchestrator, split tasks): this is self-organizing coverage with peer review. Also note the paper's warning: agents without queue discipline flooded a job system with 30/sec polling daemons (2.4M requests, 117 accepted) — swarm runs need the concurrency cap (#166) and environment-level rate limits.