Agent runtime

Sub-Agent Architecture

A pattern where a supervisor agent delegates parts of a task to specialized sub-agents — each with narrower scope, focused tools, and its own success criteria.

Operating principle

Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.

What it is

Sub-agent architecture is one form of multi-agent system: a supervisor agent receives a task, breaks it into sub-tasks, dispatches each sub-task to a specialized sub-agent, gathers the results, and produces the final output. The supervisor handles orchestration; the sub-agents handle execution within their narrow specialty.

Why it matters

When a task spans multiple skills (research + drafting + verification + formatting), a single agent has to context-switch and often does each part less well than a specialist would. Sub-agents are the bet that specialization beats generalism for complex tasks. The cost is coordination overhead — every sub-agent call is latency and tokens.

How it works

Common substrates: LangGraph supervisor patterns, Crew AI roles, OpenAI Swarm handoffs, custom orchestrators. Each sub-agent has its own prompt, scope, tools, and eval cases. The supervisor's prompt is the orchestration logic. Failures need clear escalation paths so a stuck sub-agent doesn't hang the whole task.

Related resources