Routing, delegation, and topology
The most valuable enterprise problems exceed any single agent, so advanced solutions need a team of specialized agents. Coordination is an evolution: as a system moves from single-agent (Levels 1–3) to multi-agent (Level 4+), patterns become more dynamic and decentralized. Three patterns set the structure.
Agent Router (intent-based routing)
The "Hello World" of coordination — a switchboard that decouples the user's intent from the executing agent, combining semantic intent extraction (the "what") with graph-constrained routing (the "who"). Replaces brittle hardcoded if "sales" in query logic and scales to dozens of agents.
Task Delegation Frameworks
Two approaches: the Supervisor (Orchestrator) Architecture — a central agent decomposes and delegates, giving predictable, auditable flow; and the Swarm Architecture — decentralized agents self-organize, trading predictability for adaptability.
Agent Composition Topologies
Blackboard (shared knowledge evolves on a common board), Contract-Net (market-based task allocation by bidding), and Supervision Trees (hierarchical fault tolerance).
Shared plans and shared knowledge
Agents that work together need both a shared plan and a shared pool of knowledge — plus a way to manage the tools they call.
- Multi-Agent Planning
- Decomposes a complex goal into a coordinated plan of sub-tasks distributed across agents.
- Knowledge Sharing
- Pools what agents learn into collective intelligence that improves the system over time.
- Tool Routing
- Manages which agent uses which capability, preventing conflicts over shared tools.
Key concept · Collaboration requires shared context
A team of agents is only greater than the sum of its parts when it operates on common ground: a shared plan (Multi-Agent Planning) coordinates what gets done, and a shared pool of knowledge (Knowledge Sharing) lets the collective improve. Coordination then extends to execution — Tool Routing decides who wields which capability.
Protocols for disagreement
Any distributed system generates friction — competing goals, contested resources, direct clashes. Without structured protocols, that friction becomes chaos or deadlock. Five patterns handle it gracefully.
| Pattern | What it resolves |
|---|---|
| Consensus | Agreeing on a shared fact or decision across agents. |
| Negotiation | Resolving conflicting goals between agents. |
| Resource Allocation | Distributing scarce shared resources fairly and efficiently. |
| Conflict Resolution | Handling direct clashes when actions or claims collide. |
| Formation Control | Coordinating the spatial/logical arrangement of agents. |
Key concept · Coordination maps to maturity
The choice and complexity of patterns tracks the system's maturity. Foundational multi-agent systems (Level 4) rely on centralized, predictable patterns like the Supervisor; advanced and self-correcting systems (Levels 5–6) need dynamic, decentralized patterns — Negotiation, Consensus, and meta-agents — to manage emergent behavior.
Chapter 5 quiz
Fifteen questions on routing, delegation, topologies, collaboration, and the disagreement protocols. Answer first, then expand Show answer.
Part A · Multiple choice
The Agent Router pattern combines which two mechanisms?
- A Retry and timeout
- B Semantic intent extraction and graph-constrained routing
- C Fine-tuning and RAG
- D Sandboxing and firewalls
Show answer
B. Understanding the "what" (intent) and deciding the "who" (routing) — a switchboard that scales beyond hardcoded conditionals.
Which task-delegation approach gives the most predictable, auditable control flow?
- A Swarm Architecture
- B Supervisor (Orchestrator) Architecture
- C Blackboard
- D Contract-Net
Show answer
B — Supervisor. A central orchestrator decomposes and delegates; Swarm trades predictability for adaptability.
Which composition topology uses market-based bidding to allocate tasks?
- A Blackboard
- B Contract-Net
- C Supervision Trees
- D Formation Control
Show answer
B — Contract-Net. Blackboard shares evolving knowledge; Supervision Trees provide hierarchical fault tolerance.
Which pattern resolves conflicting goals between agents?
- A Consensus
- B Negotiation
- C Tool Routing
- D Knowledge Sharing
Show answer
B — Negotiation. Consensus is for agreeing on facts; Negotiation resolves goal conflicts.
What does the Tool Routing pattern manage?
- A Which agent uses which capability/tool
- B The physical location of servers
- C Model fine-tuning schedules
- D User authentication
Show answer
A. It coordinates capability use across agents, preventing conflicts over shared tools.
Which pattern creates collective intelligence that improves the system over time?
- A Formation Control
- B Knowledge Sharing
- C Conflict Resolution
- D Agent Router
Show answer
B — Knowledge Sharing. It pools what agents learn into a shared, improving pool.
At which maturity do dynamic, decentralized patterns (Negotiation, Consensus, meta-agents) become necessary?
- A Levels 1–3
- B Level 4 only
- C Levels 5–6 (advanced/self-correcting)
- D Never
Show answer
C. Level 4 relies on centralized patterns like Supervisor; Levels 5–6 need decentralized ones to manage emergent behavior.
Part B · True or false
Coordination between agents can be assumed to emerge naturally without explicit patterns.
Show answer
False. Coordination is architected, not assumed — it requires explicit patterns for delegation, planning, and interaction.
The Blackboard topology provides a shared space where knowledge evolves collaboratively.
Show answer
True. Agents read and write to a common board as shared knowledge develops.
A Swarm Architecture centralizes all decisions in one orchestrator agent.
Show answer
False. Swarm is decentralized and self-organizing; the Supervisor is the centralized approach.
Structured protocols for disagreement help prevent deadlock in distributed agent systems.
Show answer
True. Consensus, Negotiation, Resource Allocation, and Conflict Resolution manage competition gracefully.
Part C · Short answer
Name the two task-delegation frameworks.
Show answer
Supervisor (Orchestrator) Architecture and Swarm Architecture.
Name the three agent composition topologies.
Show answer
Blackboard, Contract-Net, and Supervision Trees.
Distinguish Consensus from Negotiation.
Show answer
Consensus = agreeing on a shared fact/decision; Negotiation = resolving conflicting goals between agents.
Name the two patterns that provide "shared context" for collaboration.
Show answer
Multi-Agent Planning (a shared plan) and Knowledge Sharing (a shared pool of knowledge).