subagents

Thariq, an engineer at Anthropic, lists subagents as one of the five options a Claude Code user has at the end of each turn, next to continuing, rewind, /clear and /compact. That placement carries the concept: delegation is first a context management decision. His test is narrow, will I need this intermediate output or only the conclusion? Blocks of work that generate a lot of noise, verifying a result, reading another codebase, writing documentation, go to an agent with its own clean context window, and the main conversation stays on the high-level objective.

A second use is adversarial. Boris Cherny spawns several parallel Claude instances for code review: a first pass on style guidelines, project history and obvious bugs, then five further subagents poking holes in those findings. He reports that the combination catches real issues without false alarms. He also uses them for code migrations, where a main agent builds the to-do list and subagents take items in parallel. At Every, Dan Shipper and Kieran Klaassen place twelve parallel subagents in the assess step of Compound Engineering, each judging the code from one angle (security, performance, complexity, architecture); that step consumes roughly 40% of a cycle.

Alan Blount extends the pattern to models rather than roles: in a coding harness, subagents are mapped onto distinct models, the fast one in front, the deep planner at architecture checkpoints. Thariq expects Claude to manage these mechanisms automatically in time. For now, choosing what to delegate remains a user skill.