Subagents Conwidered Harmful

Subagents often fail because they receive a compressed assignment rather than the conversation that explains it. A summary can omit key requirements, prior corrections, or rejected approaches, leaving the worker unable to tell what it does not know. One example involved delegating research into whether a tool duplicated products from Anthropic and OpenAI: the summary omitted both companies, so the subagent investigated only the tool’s code. A cited comparison found summary-based delegation slower and more expensive than doing the task in the main session. Forking a conversation gives a subagent the full context, but can burden it with irrelevant history and tokens. The key choice is not whether a task is small, but whether it requires judgment grounded in the project’s intent and decisions.

The author argues that agents struggle to judge what a summary leaves out, because their own context makes an incomplete brief appear sufficient. Guidance should therefore focus on the worker’s missing knowledge: use a fork for holistic decisions and a fresh, focused task for mechanical work. The proposed “smart fork” selects relevant pieces of the original conversation with a fast decision model, preserving their wording while dropping irrelevant material. An outside evaluation of a related method retained 99 of 105 necessary facts using about a quarter of the full transcript’s tokens, compared with 82 facts in a written summary. The author’s own experiment looks promising but remains unproven; they continue using subagents, favoring forks over summary-only delegation.