How to actually evaluate an agent swarm

Separate single-agent and swarm metrics with budget caps.

Swarm 357 ships open-core orchestration with honest maturity labels. Read the documentation for installation, API reference, and eval methodology.

Key takeaways

• Operators own API keys, budgets, and side-effect approvals
• Security defaults: Bash gate, filesystem confinement, HITL in production
• Eval baselines report single-agent and swarm metrics separately

Get started or view latest eval baseline.