subagents

Overview

Subagents are specialized subordinate agents that a primary AI agent can delegate work to in order to break large tasks into smaller, more manageable parts. Instead of one agent carrying the full context, planning, and execution burden, subagents can take on scoped responsibilities such as exploration, implementation, summarization, or validation. In practice, they are used to improve performance on tasks that exceed model context limits, enable parallel execution, and create more controllable multi-agent workflows.

For AI Product Managers, subagents matter because they represent an emerging design pattern for building reliable agentic products. They help teams address common product constraints such as limited context windows, long-running tasks, orchestration complexity, and the need for predictable outputs. The concept is increasingly appearing across tools like Claude Code, OpenAI Codex, and Cursor, often alongside steering mechanisms such as rules, hooks, skills, and custom agent definitions.

Key Developments

• 2026-01-23 — Cursor introduced subagents for parallel task execution, highlighting faster execution, improved context usage, and support for long-running tasks.
• 2026-02-22 — Boris Cherny described using subagents with Git worktrees to parallelize large codebase migrations, assigning each agent a subset of folders while a main agent handled merge conflicts.
• 2026-03-17 — OpenAI Codex announced general availability of subagents and support for custom agents, including TOML-defined agent configurations and patterns similar to Claude Code's explorer/worker/default setup.
• 2026-03-18 — Simon Willison explained subagents as a way to manage LLM context limits by splitting large tasks into smaller agentic components while preserving top-level context.
• 2026-06-04 — A newsletter mention noted that while both subagents and workflows can handle extensive context, workflows may be a better fit for many tasks when deterministic behavior is more important.
• 2026-06-19 — Claude Code announced new steering controls including CLAUDE.md files and mechanisms such as skills, hooks, rules, and subagents to guide agent behavior.

Relevance to AI PMs

1. Design around context limits: Subagents offer a practical way to handle tasks that exceed a single model's memory window. PMs can use them to structure product flows where a top-level agent delegates research, coding, summarization, or review to smaller scoped agents.

2. Improve speed and throughput: In products that support parallelism, subagents can execute task components simultaneously. This is especially useful for code generation, migration, analysis, and long-running operations where user value depends on reducing turnaround time.

3. Balance autonomy with control: Subagents are often paired with steering systems like rules, hooks, skills, and agent configuration files. PMs should think carefully about when to use flexible agent delegation versus deterministic workflows, depending on the product's tolerance for variance, auditability, and failure modes.

Related

• claude / claude-code — Claude Code has surfaced subagents as part of a broader set of steering controls for agent behavior.
• claudemd — CLAUDE.md files help define instructions and guidance that can shape how agents and subagents operate.
• skills, hooks, rules — These are complementary control mechanisms often used alongside subagents to constrain or specialize behavior.
• openai-codex — Codex added subagents and custom agents, showing that the pattern is spreading across coding-agent platforms.
• custom-agents — Custom agents are a close cousin to subagents; they allow predefined roles, behaviors, or capabilities within an agent system.
• toml — TOML is used in some systems to define custom agent configurations.
• llm-context-limits — One of the main motivations for subagents is mitigating context-window constraints by decomposing work.
• git-worktrees — Git worktrees enable practical parallel execution for code-focused subagent setups, especially in migrations or multi-folder refactors.
• cursor — Cursor introduced subagents for parallel task execution and long-running tasks.
• workflows — Workflows and subagents overlap, but workflows are often preferred when deterministic orchestration is more important than agent flexibility.
• boris-cherny — Boris Cherny provided a concrete example of using subagents operationally in large-scale code migration work.