$ swarm command --scale 250 "audit auth system" 🐝 Hive activated · 250 agents · 15 models · 3 families consensus: 94% ✓ Cross-family validation passed ✓ Shadow score: 96/100 ✓ 3 critical findings synthesized → Final report delivered in 4m 12s
For small tasks, a single AI is fine. But for security audits, architecture reviews, and migration strategies — one model means one blind spot, one context window, one confident-sounding answer with no independent check. You need consensus from independent minds that verify each other's work.
One command. Tell the swarm what you need — security audit, code review, architecture analysis. Plain English.
Agents from Claude and GPT families compete and collaborate. Different models cross-pollinate and review each other's work.
Only findings validated across model families survive. Shadow scores gate quality. One synthesized answer emerges from the colony.
15 models, not one. Claude Opus & Sonnet. GPT-5.x series. Claude Haiku. Each brings different strengths — together they catch what any single model misses.
Different model families review each other's work. Claude checks GPT. GPT checks Claude. No echo chambers — only findings that survive independent scrutiny make the cut.
Hidden quality gates you can't game. Every agent is scored — failures ÷ total × 100 — and they don't know they're being watched. Bad work gets caught automatically.
Every commander runs in its own context window.
Different model families ensure diverse perspectives.
5 Commanders × ~50 workers each = 250 agents, each with its own context window
Multiple independent minds converge on one synthesized truth.
Thorough analysis with full cross-family validation. Architecture reviews, security audits, migration planning.
$ swarm command --scale 100 "audit security posture"Every layer of the swarm is engineered to maximize signal while minimizing spend. Here's how.
Context shrinks at every layer — 128K tokens at the Nexus compresses to just 128 tokens at each worker. Parents strip rationale, narrow file scope, and tighten constraints so children only receive the bytes they need.
A three-state FSM (Closed → Open → Half-Open) monitors every layer. If 50-60% of agents fail, the breaker trips — no new agents spawn, costs stop climbing, and a recovery probe tests before the swarm resumes.
Timeout cascade (90→60→40→30s), token ceilings per layer, output size caps, retry budgets, a concurrent-agent cap of 50, and a hard cost ceiling ($5–$20 depending on scale) that kills all agents if breached.
Agents launch in three waves — Canary (1), Probe (3), Remainder — with health gates between each. If the canary fails, the full pod never deploys. One cheap test prevents many expensive failures.
Workers use Haiku and GPT-Mini — the lightest, cheapest models. Expensive Opus and Sonnet reasoning is reserved for Commanders and the Nexus where it matters most. 60% of agents cost 10× less.
SS-50 runs $1.50–$3.50. SS-100 runs $3.50–$8. SS-250 runs $8–$16. Hard ceilings at $5, $10, and $20 guarantee you never get a surprise bill — even if every agent retries at maximum.
| Scale | Agents | Typical Cost | Hard Cap | Wall-Clock |
|---|---|---|---|---|
| SS-50 | ~36-52 | $2.50 | $5 | ~30s |
| SS-100 | ~89 | $5.50 | $10 | ~45s |
| SS-250 | ~316 | $10 | $20 | ~65–90s |
Progressive refinement: discover → validate → confirm
consensus = confidence×0.40 + evidence×0.30 + scope×0.15 + coverage×0.15 − conflict_penaltyOne command. Then type swarm command.
curl -fsSL https://raw.githubusercontent.com/DUBSOpenHub/swarm-command/main/quickstart.sh | bashRequires an active Copilot subscription