01 · Mental model
Start with ownership, not agent count
A manager pattern keeps one component responsible for combining specialist outputs. A handoff transfers control to a specialist. Code-led orchestration makes sequence, approval, and failure behavior explicit in application logic.
Use multiple agents when roles have genuinely different instructions and evidence, work can be separated, and handoffs can be typed and reviewed. Otherwise, one agent with focused tools is usually easier to secure, observe, and evaluate.
02 · Visual explanation
03 · Compare and decide
When complexity earns its place
| Decision lens | Stay single-agent | Consider multi-agent |
|---|---|---|
| Expertise | One instruction set covers the task | Roles need different evidence and rubrics |
| Context | All steps use the same information | Context can be isolated by specialist |
| Handoff | One definition of done | Typed artifacts have clear owners |
| Assurance | Self-check plus human review is enough | Independent review adds measurable value |
04 · Cybersecurity example
Cybersecurity delivery pack
A delivery pack needs architecture, compliance, detection, and project-planning expertise.
Code routes bounded work to specialists.
Each specialist returns a typed artifact.
A manager reconciles disagreements.
Deterministic gates validate the assembled pack.
Outcome: Multiple agents are justified because expertise, artifacts, and review responsibilities are separable.
05 · What to remember
The 60-second recall
Agent count is not a value metric.
Typed artifacts are safer than uncontrolled agent conversation.
Code-led orchestration is strong when approvals and failures must be explicit.
Teach-back prompt: Explain this concept to a teammate using the diagram, then name one failure mode and the control that stops it.
06 · Questions people ask
FAQ
07 · Primary sources