multi-agent-architecture-reference（多 Agent 架构参考）

Multi-Agent Architecture Reference

Step 1: Characterize the Task

Answer these four questions before selecting a topology:

1. Task independence: Can sub-tasks run in parallel without shared state? (YES → Swarm or Fan-out)
2. Task types known: Is the set of task types stable and deterministic at design time? (YES → Supervisor)
3. Phase complexity: Does the work require multi-stage sub-orchestration? (YES → Hierarchical or Conductor)
4. Stakes: Does an incorrect outcome require multi-reviewer agreement? (YES → Consensus Voting)

Step 2: Apply the Topology Decision Matrix

Token costs are relative to single-agent baseline (as of 2026). Use as order-of-magnitude guidance.

Step 3: Check Failure Mode Taxonomy

Before finalizing topology, verify mitigation for relevant failure modes:

SE-M01: Coordinator Overload

• Topologies affected: Supervisor, Conductor, Hierarchical root
• Symptom: Single coordinator receives more traffic than it can route
• Fix: Distribute coordination or add routing replicas; use wave-executor for fan-out

SE-M02: Swarm Deadlock

• Topologies affected: Swarm, Consensus Voting
• Symptom: Agents wait for each other's consensus indefinitely
• Fix: Timeout + majority-vote with tie-breaker; set consensus_timeout_ms

SE-M03: Cascade Failure

• Topologies affected: Hierarchical
• Symptom: A mid-level agent failure halts all downstream agents
• Fix: Circuit breakers at each tier; retry with backoff; fallback agents

SE-M04: Token Runaway

• Topologies affected: Hierarchical
• Symptom: Spawning too many levels burns tokens exponentially
• Fix: Set max_depth=3; monitor token budget per level; prefer Conductor over deep Hierarchical

SE-M05: Orphaned Tasks

• Topologies affected: Swarm
• Symptom: Agents drop tasks when no ownership is clear
• Fix: Assign task IDs; use TaskUpdate tracking; require TaskUpdate(in_progress) on pickup

Step 4: Apply Escalation Path

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)

Step 5: Reference Existing agent-studio Patterns

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.

Example 1: Code Review Pipeline

• Task: Review 5 files for security, quality, and style
• Character: Tasks are independent (YES), parallel OK (YES)
• Topology: Fan-out/Fan-in (~8x)
• Pattern: wave-executor skill — spawn 3 reviewers in parallel, aggregate results

Example 2: Feature Implementation

• Task: Design → Implement → Test → Document
• Character: Sequential phases, ordered steps (YES)
• Topology: Conductor (~6x)
• Pattern: master-orchestrator with TaskUpdate coordination between phases

Example 3: Architecture Decision

• Task: Choose between 3 database options for production system
• Character: High stakes, requires agreement (YES)
• Topology: Consensus Voting (~12x)
• Pattern: consensus-voting skill — 3 architect agents vote, majority decides

Example 4: Batch Agent Creation

• Task: Create 10 new agents from specs
• Character: Independent tasks (YES), fault tolerance > ordering (YES)
• Topology: Swarm (~8x)
• Pattern: swarm-coordination skill with task ID assignment per agent

<best_practices>

• Default to Conductor (master-orchestrator) — it is the lowest-risk pattern for most tasks
• Never use Hierarchical beyond depth=3 (token runaway risk SE-M04)
• Always assign TaskUpdate(in_progress) on task pickup in Swarm to prevent SE-M05
• Use Fan-out (wave-executor) instead of Swarm when tasks have clear aggregation boundary
• Add consensus gate only for genuinely high-stakes decisions — 12x token cost is significant
• Document token budget per topology tier when spawning Hierarchical
• Cross-reference failure mode taxonomy before finalizing topology choice </best_practices>

Iron Laws

1. ALWAYS start with Conductor — default to master-orchestrator for MEDIUM/HIGH tasks; only escalate to Hierarchical when sub-orchestration is explicitly required by the task structure.
2. NEVER exceed depth=3 in Hierarchical — token cost grows exponentially at each tier; depth >3 triggers SE-M04 (token runaway) and is considered an architectural defect.
3. ALWAYS assign TaskUpdate(in_progress) on Swarm task pickup — missing task ownership is the root cause of SE-M05 (orphaned tasks); every agent in a swarm must call TaskUpdate before doing work.
4. NEVER use Consensus Voting for low-stakes decisions — 12x token multiplier is justified only for architecture decisions, security approvals, or irreversible production changes.
5. ALWAYS cross-reference the failure mode taxonomy before finalizing topology — each topology has documented failure modes (SE-M01 through SE-M05); skipping this review leads to production incidents.

Anti-Patterns

Memory Protocol (MANDATORY)

Before starting:

Read .claude/context/memory/learnings.md to check for prior multi-agent architecture decisions.

After completing:

• New topology decision → Append to .claude/context/memory/decisions.md
• Failure mode encountered → Append to .claude/context/memory/issues.md
• New pattern discovered → Append to .claude/context/memory/learnings.md

ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.

Related Skills

• wave-executor — Fan-out/Fan-in implementation
• swarm-coordination — Swarm topology execution
• consensus-voting — Byzantine consensus for high-stakes decisions
• architecture-review — Validate topology choices against NFRs
• complexity-assessment — Determine complexity level before topology selection