subagents

Overview

Subagents are specialized subordinate agents that take on a portion of a larger task under the direction of a primary agent or orchestration layer. Instead of forcing a single model run to hold all instructions, files, and intermediate reasoning in one context window, subagents let teams decompose work into smaller, bounded units such as exploration, implementation, review, migration, or synthesis. In practice, they are used to parallelize work, preserve top-level context, and improve reliability on tasks that exceed a single model’s effective memory limits.

For AI Product Managers, subagents matter because they represent a practical design pattern for building scalable agentic systems. They show up in coding tools like Claude Code, OpenAI Codex, and Cursor, and are increasingly paired with steering controls such as rules, hooks, skills, and project instruction files. PMs evaluating agent products need to understand when subagents improve throughput and context handling, when deterministic workflows are a better fit, and how to design user experiences, governance, and system architecture around multi-agent task delegation.

Key Developments

• 2026-01-23 — Cursor introduced subagents for parallel task execution, emphasizing faster execution, improved context usage, and support for long-running tasks.
• 2026-02-22 — Boris Cherny highlighted a practical pattern for using subagents with Git worktrees to parallelize large codebase migrations, with a main agent coordinating results and resolving 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 described subagents as a solution to LLM context limits, showing how splitting larger tasks into smaller agentic components can preserve top-level context and improve outcomes.
• 2026-06-04 — A newsletter mention noted that while both subagents and workflows can manage extensive context, deterministic workflows may be a better fit for many tasks depending on reliability needs.
• 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. Designing around context limits
Subagents give PMs a concrete way to scope products for tasks that exceed a single model’s context window. Instead of asking one agent to do everything, product teams can decompose tasks into specialized sub-jobs with clear inputs and outputs.

2. Improving speed and throughput
For coding, research, migration, and analysis products, subagents can execute work in parallel. PMs can use this pattern to reduce time-to-result, especially for large repositories, multi-file edits, or broad information gathering.

3. Choosing between autonomy and determinism
Subagents are powerful, but they also add orchestration complexity. PMs need to decide when open-ended delegation is appropriate versus when workflow-based, deterministic systems are safer, more debuggable, and easier to govern.

Related

• Claude / Claude Code — Claude Code is one of the main contexts where subagents were discussed, especially as part of broader steering controls.
• CLAUDE.md — Project-level instruction files that can help shape how subagents behave within Claude Code.
• Skills, hooks, rules — Complementary steering mechanisms that guide agent behavior, often working alongside subagents.
• LLM context limits — A core reason subagents exist; they help partition work so no single agent must hold all context at once.
• OpenAI Codex — Added subagents and custom agents, showing that the pattern is becoming cross-platform.
• Custom agents — Closely related to subagents; teams often define specialized roles for different parts of a task.
• TOML — Used in some systems to define custom agent configurations.
• Git worktrees — Useful infrastructure for parallel subagent execution in software engineering workflows.
• Cursor — Another tool that introduced subagents for parallel execution and long-running tasks.
• Workflows — Often contrasted with subagents; workflows offer more deterministic orchestration.
• Boris Cherny — Shared a notable applied pattern for using subagents in large migrations.
• claudemd — Related naming/reference variant tied to CLAUDE.md-style steering.