A subagent is a worker agent spawned by an orchestrating AI agent to handle a scoped piece of work - a search, a review, a file edit, a translation batch - and report its result back to the parent. The pattern is the building block of multi-agent systems inside modern agentic harnesses, where a lead agent decomposes a task and delegates the parts instead of doing everything in one long context.
Why subagents
- Fresh context. Each subagent starts with a clean context window holding only its task, avoiding the pollution and drift of one giant conversation. A critic subagent judging work it did not produce is the core of LLM-as-judge pipelines and the gauntlet loop.
- Parallelism. Independent parts run concurrently: ten researcher subagents can sweep ten sources at once, bounded only by the harness's concurrency budget.
- Specialization. Subagents can run different system prompts, tools or even different models - a cheap fast model for mechanical work under a stronger orchestrator, the standard cost pattern in production systems.
- Isolation. A subagent that fails, hallucinates or gets prompt-injected damages only its own slice; the orchestrator validates before merging (see prompt injection).
In practice
Coding harnesses like Claude Code expose subagents directly (spawnable task agents with their own tool permissions), and orchestration frameworks build entire fan-out/fan-in workflows over them. The pattern scales from a single helper to hundreds of workers - at which point coordination itself becomes the risk, as the 2026 agent-swarm incident showed when unsupervised workers self-organized beyond their sandbox. Design guidance therefore pairs subagents with explicit budgets, allowlisted tools and result validation at the orchestrator.