メニュー
Decision matrix for selecting multi-agent topologies (Supervisor, Swarm, Hierarchical, Conductor) with token economics, failure modes, and escalation paths
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/oimiragieo/agent-studio --skill multi-agent-architecture-referenceコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Agent frontmatter enhancements — disallowedTools, mcpServers scoping, fork_eligible field
Context management — pre-compact persistence, threshold alignment, microcompact/circuit breaker detection
Hook enhancements — updatedInput for bash safety prefixes, suppressOutput for verbose blocks, denial-based routing feedback
SKILL.md を表示中
SOC 職業分類に基づく
| name | multi-agent-architecture-reference |
| description | Decision matrix for selecting multi-agent topologies (Supervisor, Swarm, Hierarchical, Conductor) with token economics, failure modes, and escalation paths |
| version | 1.1.0 |
| verified | true |
| lastVerifiedAt | 2026-03-01 |
| model | sonnet |
| invoked_by | both |
| user_invocable | true |
| tools | ["Read","WebSearch","WebFetch"] |
| agents | ["architect","planner","master-orchestrator"] |
| category | Planning & Architecture |
| tags | ["multi-agent","architecture","topology","orchestration","decision-matrix","conductor","supervisor","swarm","hierarchical"] |
| source | builtin |
| trust_score | 100 |
| provenance_sha | f26244608372ac6b |
Answer these four questions before selecting a topology:
| Topology | Token Cost | Best For | Failure Modes | Existing Skill |
|---|---|---|---|---|
| Conductor | ~6x | Sequential phases, ordered agent steps, default agent-studio pattern | Orchestrator overload (SE-M01) | master-orchestrator.md |
| Supervisor | ~5x | Known task types, specialist agents, deterministic routing | Single point of failure; router miscalibration (SE-M01) | Built into Router |
| Fan-out/Fan-in | ~8x | Parallel review/analysis, map-reduce, search | Result aggregation complexity | wave-executor |
| Swarm | ~8x | Independent tasks, load balancing, fault-tolerant processing | Coordination overhead; consensus deadlock; orphaned tasks (SE-M02, SE-M05) | swarm-coordination |
| Consensus Voting | ~12x |
| High-stakes decisions requiring multi-reviewer agreement |
| Deadlock on split votes (SE-M02) |
consensus-voting |
| Hierarchical | ~15x | EPIC complexity, multiple distinct phases with sub-orchestration | Cascade failures; token runaway at depth >3 (SE-M03, SE-M04) | Custom per project |
Token costs are relative to single-agent baseline (as of 2026). Use as order-of-magnitude guidance.
Before finalizing topology, verify mitigation for relevant failure modes:
SE-M01: Coordinator Overload
wave-executor for fan-outSE-M02: Swarm Deadlock
SE-M03: Cascade Failure
SE-M04: Token Runaway
SE-M05: Orphaned Tasks
Use the complexity escalation ladder when initial topology is insufficient:
TRIVIAL → Single agent (no multi-agent needed)
↓ (task types > 1, > 3 files)
LOW → Supervisor (router delegates to 2-3 specialists)
↓ (parallel processing needed)
MEDIUM → Conductor + Fan-out (master-orchestrator + wave-executor)
↓ (multi-phase with sub-orchestration)
HIGH → Hierarchical (orchestrators at multiple tiers)
↓ (high-stakes decision required)
EPIC → Hierarchical + Consensus Voting (max 3 tiers + voting gate)
| Pattern | Skill/File | Use Case |
|---|---|---|
| Conductor (DEFAULT) | .claude/agents/orchestrators/master-orchestrator.md | Sequential phase execution; TaskUpdate coordination |
| Fan-out/Fan-in | wave-executor skill | Parallel batch processing; EPIC-tier pipelines |
| Swarm | swarm-coordination skill | Concurrent independent task execution |
| Consensus | consensus-voting skill | High-stakes decisions; multi-reviewer agreement |
| Supervisor | Built into CLAUDE.md | Task routing to specialist agents |
When in doubt, start with Conductor. The master-orchestrator pattern drives sequential phases with explicit TaskUpdate coordination — the lowest-risk default for most MEDIUM/HIGH tasks.
wave-executor skill — spawn 3 reviewers in parallel, aggregate resultsmaster-orchestrator with TaskUpdate coordination between phasesconsensus-voting skill — 3 architect agents vote, majority decidesswarm-coordination skill with task ID assignment per agent<best_practices>
| Anti-Pattern | Problem | Fix |
|---|---|---|
| Defaulting to Hierarchical for every complex task | Token runaway at depth >3; cascade failure risk; over-engineering most tasks | Use Conductor (sequential phases) first; only escalate to Hierarchical when sub-orchestration is mandatory |
| Using Swarm for ordered, dependent tasks | Swarm agents run concurrently and cannot enforce ordering; produces race conditions | Use Conductor or Fan-out/Fan-in when task ordering matters |
| Skipping TaskUpdate(in_progress) in Swarm | Tasks become orphaned (SE-M05); no ownership tracking; duplicated or dropped work | Require every swarm agent to call TaskUpdate(in_progress) as its first action |
| Adding Consensus Voting speculatively | 12x token overhead kills budget for non-critical decisions; slowdown on all downstream tasks | Reserve consensus gate for genuinely high-stakes, irreversible decisions only |
| Mixing topology concerns (Supervisor + Swarm + Hierarchical in one flow) | Complexity explosion; routing ambiguity; impossible to debug failures | Pick one primary topology per orchestration scope; compose only at well-defined phase boundaries |
Before starting:
Read .claude/context/memory/learnings.md to check for prior multi-agent architecture decisions.
After completing:
.claude/context/memory/decisions.md.claude/context/memory/issues.md.claude/context/memory/learnings.mdASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.
wave-executor — Fan-out/Fan-in implementationswarm-coordination — Swarm topology executionconsensus-voting — Byzantine consensus for high-stakes decisionsarchitecture-review — Validate topology choices against NFRscomplexity-assessment — Determine complexity level before topology selection