swarm-mesh

Use when multiple perspectives are needed on the same work — every agent reviews every other agent’s output. Also use when security-critical code, complex algorithms, or multi-domain changes need peer review from multiple angles.

Model	Source
opus	pack: swarms

Full Reference

Swarm — Mesh Topology

Full peer review. Every agent reviews every other agent’s work. Consensus required before any output is accepted. High-trust, high-quality — at the cost of O(N²) reviews. Hard limit: N ≤ 4. More agents means exponentially more reviews; beyond 4, switch to hierarchical with selective review.

---

Quick Reference

Item	Value
Pattern	Every agent reviews every other agent
Review count	N × (N-1)
Hard limit	N ≤ 4
Phases	Phase 1: parallel work → Phase 2: cross-review
Revision cycles	Max 2, then escalate
Speed	Slowest topology — use only when quality demands it

---

Reference Index

I want to…	File
Understand the topology diagram and review matrix	reference/topology-and-review-matrix.md
Implement Phase 1 dispatch and Phase 2 cross-review	reference/implementation.md
Know when to use mesh vs other topologies	reference/when-to-use.md

Usage: Read the reference file matching your current task from the index above. Each file is self-contained with code examples and inline gotchas.

---

Agent Teams

For heavy multi-domain tasks, mesh agents can be organized into Agent Teams — small groups of 3-5 agents that collaborate with a shared collaboration_mode before cross-reviewing each other’s output.

When to use Agent Teams:

• Task is heavy weight (5+ files, multi-domain, parallelizable)
• Work spans domains that benefit from both specialization and peer review
• You need consensus across perspectives before accepting output

Team size: 3-5 agents. Below 3 is just peer review. Above 5 is exponential review overhead — switch to hierarchical.

Configuration:

{
  "collaboration_mode": "peer_review",
  "agents": [
    { "role": "security-reviewer", "focus": "auth flows and input validation" },
    { "role": "architect", "focus": "data model and API design" },
    { "role": "implementer", "focus": "code correctness and test coverage" }
  ]
}

Flow:

1. Phase 1 — each agent works independently on assigned domain
2. Phase 2 — every agent reviews every other agent’s output (collaboration_mode: peer_review)
3. Consensus — all agents must agree before output is accepted
4. Max 2 revision cycles, then escalate to hierarchical

---

Announcement

┏━ ⚡ swarm-mesh ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ [one-line description of what’s being peer-reviewed] ┃ ┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛