Subagents: Great Idea, Terrible Group Project

Claude Code and Codex are both pushing you to split work across a squad of AI agents. Before you spin up five of them for a login page, read this.

The Hook: You ask one agent to build a feature, it spawns a few helpers, and suddenly your main chat is a relay race where you re-explain the same backstory to everyone. Not exactly the productivity glow-up you were promised.

The Meat: Spotted on a very candid thread: the re-explaining happens because every subagent starts with a blank slate, so the main chat ends up being the map. Commenters had a few fixes:

• Put the context in files, not the chat. As one user neatly summed it up, "only if the shared decisions live outside the chat, otherwise it's just re-briefing N times." A plan or spec file in the repo that every agent reads first means the handoff is just the task.
• Use CLAUDE.md for the common stuff. Stack, conventions, and test commands go there. Per the thread, agent-team teammates load it automatically, so spawn prompts stay short. For research-only agents, one commenter noted you can set omitClaudeMd: true in a custom subagent's frontmatter to keep them light.
• Split by concern, not by feature. "Build the feature" is the airplane. A subagent should get the wing. If two agents need the same backstory, it's probably one task, not two.
• Mind the leash. Several people warned that letting 3 or 4 subagents loose unsupervised mostly burns your token quota. One said a single capable model often beats a spiral of agents confusing each other.

Also spotted: a few replies pitching shared-memory tools (Engram, AtlasMind). The idea is the same as the repo-file fix, so treat those as optional extras, not requirements.

The Takeaway: Parallel agents pay off when the pieces are truly independent (separate files, a test run) and each has its own finish line. For tightly coupled work, one agent is usually faster. Write the shared brief once, point everyone at it, and only then hit "split."