/agentsop-agent-topology-selection

name: agentsop-agent-topology-selection
version: 0.1.0
description: >-
  Cross-framework enhancement overlay for choosing a multi-agent topology BEFORE writing any
  agent. A binary-question rubric — is single-agent + tools enough? do agents need to know
  about each other? does the output need one voice? — maps the answer to single-agent /
  supervisor / swarm / sequential / hierarchical. Activates when a coder agent is tempted to
  "split the work into roles" or reaches for a multi-agent framework. Encodes the *selection
  rubric* that the per-framework skills assume but never surface. Search keywords: when to
  use multi-agent, single vs multi agent, do I need multiple agents, supervisor vs swarm,
  multi-agent vs single agent, agent team design.
overlay: true
cross_links: [crewai, langgraph, bounded-loop]

Multi-Agent Topology Selection · SOP (ENHANCE overlay)

> Overlay posture: this skill decides *whether and which* topology. It does not > teach the API — descend to `[[crewai]]` or `[[agentsop-langgraph]]` for that. Every > load-bearing claim carries an inline source tag resolving in > `references/R1-source-evidence.md`.

---

1. 何时激活 (When to Activate)

Activate when **any** of the following fire:

• The task description contains "team of agents", "researcher + writer + reviewer",

"manager agent", "agents that hand off", "split this into roles", or "multi-agent".

• A coder agent is about to instantiate ≥2 agents (CrewAI `Agent(...)` × N,

LangGraph supervisor/swarm, OpenAI Swarm handoffs) and has **not yet** justified why a single agent with tools is insufficient.

• Someone is choosing between CrewAI `Process.sequential` vs `Process.hierarchical`,

or LangGraph supervisor vs swarm vs hierarchical-teams, and wants the *rubric*, not the syntax.

• A multi-agent system is over budget on tokens/latency and the question is "can we

collapse agents back into one?".

Do **not** activate for: a single LLM call, a one-shot RAG query, or a fixed tool-call pipeline with no role separation. Those are the single-agent baseline this skill defends.

> Mental check: *"An agent needs agency, otherwise it's just another script."* > — João Moura, CrewAI founder `[[crewai · §1.3]]`. If you can write the control > flow in `if/else`, you do not need multiple agents — you need one agent (or a > graph) with explicit edges.

---

2. 核心心智模型 (Core Mental Model)

**Most "multi-agent" problems are single-agent + tools.** Add agents only when *context isolation* or *parallel expertise* genuinely demands it.

> "Single-agent is right for approximately 80% of cases; the trap is reaching for > multi-agent because it sounds more capable." `[[crewai · DC-1]]`

Two — and only two — forces justify a second agent:

1. **Context isolation.** One agent's working context would pollute another's (a critic that must not see its own draft's rationalisations; a tool-heavy sub-task whose 40 intermediate tool calls should not bloat the main thread). Splitting gives each agent a clean, bounded prompt. 2. **Parallel expertise.** Two *genuinely different* skills run concurrently or in strict sequence (research → write → review), where a single prompt provably cannot hold both jobs without quality collapse `[[crewai · DC-1]]`.

If neither force is present, **a single agent with the union of tools wins** — fewer hops, fewer tokens, no handoff failures. This is the baseline the rubric must beat, not the default to escape.

The selection rubric (three binary questions)

Q0  Is single-agent + tools enough?
      (no context-isolation need, no parallel-expertise need)
        YES → single-agent + tools. STOP. Do not add agents.
        NO  → ↓

Q1  Do the agents need to KNOW ABOUT EACH OTHER (peer handoff)?
        NO  → one funnels through a coordinator → SUPERVISOR
              (or static order → SEQUENTIAL, if order is fixed)
        YES → ↓

Q2  Must the OUTPUT speak with ONE VOICE / single audit funnel?
        YES → SUPERVISOR (single user-facing persona, one funnel)
        NO  → SWARM (dynamic peer handoff, last-active agent remembered)

Scaling override: ≥6 specialists that group into teams → HIERARCHICAL
(supervisor-of-supervisors). Use only for grouping, not for routing.

The two questions that actually separate the patterns: **(a) can sub-agents know each other, (b) is one user-facing voice mandated.** Everything else is tuning `[[langgraph · Case 2]]`.

---

3. SOP 工作流 (Selection Protocol)

Walk top-down. Each gate can send you *back down* the ladder — collapsing agents is as valid an answer as adding them.

Step 1 · Defend the single-agent baseline first

Ask Q0. Enumerate the would-be roles. For each, ask: *would merging it into one agent's prompt + toolset actually degrade output?* If you cannot point to a concrete failure mode (style drift, missed checklist, context bloat, parallel latency), the honest answer is single-agent + tools. Exit here ~80% of the time `[[crewai · DC-1]]`.

Step 2 · If splitting, decide static vs dynamic routing

• **Order is fixed and known at design time** (research always precedes write

precedes review) → **SEQUENTIAL**. Cheapest, most debuggable, 1× token baseline `[[crewai · §2.3]]`. In CrewAI this is `Process.sequential`; in LangGraph it is static edges A→B→C.

• **Routing must be decided at runtime** (which specialist handles *this* query) →

you need a coordinator or peer handoff → continue to Step 3.

Step 3 · Coordinator (supervisor) vs peers (swarm)

Ask Q1 then Q2.

• Agents that do **not** know each other and funnel through one router →

**SUPERVISOR**. Sub-agents are effectively tools the supervisor calls; the supervisor "translates" their output back to the user — which is *exactly* why it costs the most tokens `[[langgraph · Step 4]]`.

• Agents that **do** know each other and **no** single voice is mandated →

**SWARM**. Dynamic handoff, last-active agent stays active across turns, no translation step →