# Multi-Agent Composition Without the Hype

**Track:** Interoperability, Safety & Production — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (subagents / delegation)
**Level:** Advanced
**Prerequisites:** The Agent Loop, Workflows vs. Agents, Durable Tasks, Queues & Background Agents

**In one line:** Create another agent only when separate context, expertise, authority, or lifecycle earns the added coordination cost.

## Theory, aesthetics & inspiration

Named specialists — researcher, critic, coder, manager — feel intuitive, which is why multi-agent systems are chronically overbuilt. Every additional agent is another context that can diverge, another hop of latency and cost, another evaluation problem; Cognition's "Don't Build Multi-Agents" (2025) argued from production experience that fragile hand-offs sink most such architectures, and Anthropic's account of its multi-agent research system (2025) reported the bill even when it works — around fifteen times the tokens of an ordinary chat interaction, against roughly four for a single agent, justified only because research parallelizes cleanly. So the default is one well-equipped agent plus deterministic tools, and a second agent is earned by isolation with architectural value: different instructions or model, a narrower tool set, separate memory, different permissions, an independently durable lifecycle. Compare three decompositions — one agent with many tools, a workflow containing several model calls, multiple collaborating agents — and choose whichever yields the clearest contracts and the easiest verification. Composition should follow responsibility and authority, not anthropomorphic storytelling.

**Builder question:** What architectural boundary does this second agent enforce that a function, workflow step, or ordinary tool would not?