Agents & Subagents

Subagents are isolated model-and-tool loops delegated by a primary agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose. Their value is not simply parallelism. They protect the main context from noisy exploration, let a role use narrower tools and instructions, and create a reviewable boundary around one unit of work.

Claude Code, Codex, and Copilot all document built-in or custom subagents. The differences that matter are configuration scope, nesting, communication, result handoff, model choice, and how much of the child run the operator can inspect.

Use a Subagent When

• The task has a bounded objective and can return a concise artifact.
• Exploration or test output would pollute the main decision context.
• A specialist needs different tools, instructions, or model economics.
• Independent reviewers can reduce anchoring on one hypothesis.

Do not delegate a tightly sequential step whose output must be reasoned over continuously by the main agent. The handoff cost can exceed the context saved.

The Control Model

Treat the parent as coordinator, not as an infallible authority. Define the child contract explicitly: objective, allowed scope, inputs, output shape, verification, and stop condition. If agents can write to the same checkout, allocate file ownership or isolate worktrees before they run.

Claude Code supports custom subagent definitions and nested delegation. Codex surfaces coordinator-managed subagent threads across local clients. Copilot custom agents run as subagents with role-specific toolsets. Those are documented capabilities; they do not establish equivalent runtime quality, reliability, or cost.

The deeper lenses retain Claude Code configuration examples. The delegation mechanics transfer; product syntax and nesting limits do not.