subagents

Overview

Subagents are specialized subordinate agents that a primary AI agent can delegate work to in order to break complex tasks into smaller, more manageable components. In practice, they are used to isolate context, parallelize execution, and assign distinct roles—such as exploration, implementation, review, or migration—so the top-level agent can focus on orchestration rather than holding every detail in memory at once.

For AI Product Managers, subagents matter because they represent a practical design pattern for building more capable agentic products under real-world constraints like LLM context limits, latency, determinism, and long-running task complexity. Newsletter coverage connects subagents to coding tools such as Claude Code, OpenAI Codex, and Cursor, where they are increasingly paired with steering controls, custom agent definitions, workflows, and repository-level guidance files. This makes subagents both a product capability and a systems design choice: they can improve user outcomes, but they also introduce orchestration, governance, and UX tradeoffs that PMs need to manage carefully.

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 conflict resolution.
• 2026-03-17: OpenAI Codex announced general availability of subagents and support for custom agents, including TOML-defined agent configurations and role-based patterns similar to Claude Code.
• 2026-03-18: Simon Willison explained subagents as a strategy for handling LLM context limits by splitting larger tasks into smaller agentic components while preserving top-level context.
• 2026-06-04: Commentary in the newsletter noted that while both subagents and workflows can manage extensive context, deterministic workflows may be a better fit for many tasks than open-ended subagent orchestration.
• 2026-06-19: Claude Code announced new steering controls including CLAUDE.md files, skills, hooks, rules, and subagents, positioning subagents as part of a broader framework for guiding agent behavior.

Relevance to AI PMs

• Design around context and reliability constraints: Subagents are a concrete way to manage LLM context limits by decomposing large jobs into scoped tasks. PMs can use this pattern to improve performance on complex workflows like code migration, research synthesis, or multi-step automation.
• Choose the right orchestration model: The newsletter discussion comparing subagents and workflows highlights an important product decision: use subagents when flexibility and adaptive reasoning matter, and prefer workflows when repeatability, auditability, and deterministic outputs are more important.
• Define governance and UX boundaries: As subagents become part of products like Claude Code and Codex, PMs need to specify how users configure them, how permissions and instructions are inherited, how progress is surfaced, and how failures or conflicts are resolved across delegated tasks.

Related

• Claude / Claude Code: Claude Code is one of the main contexts where subagents were discussed, especially as part of steering controls for agent behavior.
• LLM context limits: A core motivation for subagents is reducing pressure on model memory by splitting work into smaller contexts.
• OpenAI Codex: Codex expanded subagent support and paired it with custom agents, showing the pattern is becoming cross-platform.
• Custom agents: Subagents often overlap with or are configured as custom role-based agents for specific responsibilities.
• TOML: Mentioned in connection with Codex custom agent definitions, indicating how subagents may be configured declaratively.
• Git worktrees: A practical implementation detail for parallel subagent execution in software engineering tasks.
• Boris Cherny: Shared a concrete operating pattern for using subagents to accelerate codebase migrations.
• Cursor: Early product example of subagents for parallel execution and long-running tasks.
• Workflows: Frequently compared with subagents as an alternative orchestration approach that offers more deterministic execution.
• CLAUDE.md, skills, hooks, rules: These are adjacent steering mechanisms in Claude Code that help shape how subagents behave and what instructions they follow.