subagents

Overview

Subagents are a workflow pattern in which a primary AI system delegates parts of a larger task to helper agents, often running them in parallel. Instead of forcing one model instance to hold every detail, the main agent can break work into smaller, scoped jobs—such as research, code changes, validation, or exploration—and then combine the outputs into a final result.

For AI Product Managers, subagents matter because they offer a practical way to improve speed, handle long-running workflows, and work around LLM context limits. They are increasingly showing up in coding tools and agent platforms because they help preserve top-level context while distributing specialized or isolated work to separate agents. This makes them relevant not just as a technical implementation detail, but as a product capability that affects UX, reliability, throughput, and the kinds of tasks AI products can successfully complete.

Key Developments

• 2026-01-23 — Cursor introduced subagents for parallel task execution, highlighting faster execution, better 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, with patterns similar to Claude Code's agent setup and TOML-defined custom agents.
• 2026-03-18 — Simon Willison explained subagents as a way to manage LLM context limits by splitting large tasks into smaller agentic components, preserving top-level context and improving results on tasks that exceed model memory constraints.

Relevance to AI PMs

• Design around context limits: Subagents let PMs turn a single brittle workflow into a coordinated system where each agent handles a narrower context window. This is useful when building products that must analyze large repositories, long documents, or multi-step business processes.
• Improve speed and responsiveness: Parallel helper agents can reduce end-to-end task time by working on independent subtasks simultaneously. PMs can use this pattern to improve perceived product performance, especially in coding, research, and transformation workflows.
• Enable more reliable long-running work: For tasks like migrations, audits, or complex content generation, subagents can isolate failures and simplify retries. PMs should think about orchestration, progress tracking, merge logic, and human review points when productizing these workflows.

Related

• claude / claude-code — Claude Code helped popularize named subagent patterns such as explorer, worker, and default, making the concept concrete for developer-facing agent workflows.
• openai-codex — Codex expanded the pattern with general availability of subagents and custom agents, signaling broader platform adoption.
• custom-agents — Custom agents are closely related because subagent systems often depend on role-specific agent definitions, tools, and prompts.
• llm-context-limits — One of the main reasons subagents are useful is that they help partition work to reduce pressure on a single model's context window.
• toml — TOML is relevant because some agent systems define custom agents and their configuration in TOML files.
• git-worktrees — Git worktrees complement subagents in engineering workflows by giving parallel agents isolated working environments for code changes.
• boris-cherny — Boris Cherny's example showed a practical, high-leverage use case for subagents in large-scale code migrations.
• cursor — Cursor's product rollout highlighted subagents as a user-facing feature for faster, parallel task execution.