Module 08 · intermediate · 1 hour

Subagents

Subagents help when a task can be divided into independent, well-bounded responsibilities. Each agent needs a clear objective, scope, output contract, and tool posture.

01

Define a specialist

Subagent roles are configured with a name, description, instructions, model or reasoning preferences where supported, and tool access. The description should make delegation predictable.

[agents.security_reviewer]
description = "Review an authorized diff for security regressions"
config_file = "agents/security-reviewer.toml"
02

Delegate bounded work

Give one concrete subtask with explicit inputs and a deliverable that can be merged into the parent task. Avoid sending two agents to edit the same file unless coordination is part of the design.

  • Research: Independent evidence
  • Implement: Separate file ownership
  • Verify: Fresh validation lane
03

Coordinate parallel agents

Parallelism helps only when the lanes are independent enough to avoid serial rework. The parent agent remains responsible for reconciling contradictions and validating the combined result.

Research docs ─┐
Inspect code ───┼─> Parent synthesis ─> Verification
Review tests ───┘
04

Know when not to delegate

Keep tightly coupled edits, tiny tasks, and work requiring constant shared state in one agent. Delegation has context and coordination cost.

Continue learning

Use the official OpenAI Codex documentation as the final reference for current commands, settings, and availability.

Return to the full curriculum.

Search Codex Me

navigateEnter openEsc close