Decision Tree: Choosing a Topology
Use this tree to pick the right agent team topology for your task.
Start here: What are you trying to do?
Understand — explore, investigate, map
- Multiple independent questions? → Parallel Explorers
- Ambiguous bug or unclear root cause? → Competing Hypotheses
- One focused question? → Single subagent (no team needed)
Build — implement, deliver
- Work splits by layer/component into different files? → Feature Pod
- Large backlog of small independent items? → Task Queue
- Lead should only coordinate, never code? → Orchestrator-Only
Review — audit, critique, validate
- Multiple review perspectives needed? → Review Board
- Single perspective? → Single subagent with a focused checklist (no team needed)
Change something risky — refactor, migrate, security
- Expensive to get wrong? → Risky Refactor
- Standard risk? → Pick the topology that matches the work shape above
Any topology + quality enforcement
Quality-Gated is not a standalone topology — it layers on top of any other pattern. Apply it when you need enforced “Definition of Done” criteria.
When to overlay Quality-Gated:
- Teammates tend to mark tasks “done” prematurely
- You need automated checks before accepting work
- The cost of rework exceeds the cost of the gate
Hook example:
{
"hooks": {
"TaskCompleted": [
{
"command": "bash -c 'cd $PROJECT_DIR && npm test'",
"on_failure": "block"
}
]
}
}
Decision factors
Independence test
The most important filter. If tasks require constant back-and-forth or touch the same files, agent teams create more problems than they solve. Restructure the decomposition until workers can operate independently.
Context budget test
Large codebases, verbose logs, and multi-module exploration can exhaust a single context window. Agent teams give each worker its own full context. Subagents are cheaper but return only summaries.
Risk level
High-risk changes (security, data migrations, core architecture) benefit from independent review and plan-before-execute workflows. Risky Refactor enforces plan approval before any code changes happen.
Common combinations
These combinations are examples of topology composition – chaining or nesting primitives. See Composing Topologies for full recipes with spawn prompts.
| Scenario | Primary | Overlay |
|---|---|---|
| Ship a full-stack feature with quality checks | Feature Pod | + Quality-Gated |
| Investigate a production bug then fix it | Competing Hypotheses → Feature Pod | + Quality-Gated |
| Review a large PR from multiple angles | Review Board | — |
| Process 50 small migration tasks | Task Queue | + Quality-Gated |
| Explore unfamiliar codebase before planning | Parallel Explorers | — |