---
title: "Fold the subagents back in, keep the second opinion"
description: "Most subagents should fold back into one loop as skills, but review by a different model is the one split worth keeping."
url: "https://michael.zazon.us/notes/fold-the-subagents-back-in-keep-the-second-opinion"
published: "2026-10-28T12:00:00-04:00"
kind: "take"
subject: "ai-agents"
tags: ["agents","claude"]
---
They've been singing that song for a while. Cut down your CLAUDE.md. You don't need skills. Use fewer subagents. I'm not saying I don't believe them, but cherry picking examples to "let our big model handle everything" is also a self-interested thing to say.

It's fair to be skeptical when a vendor of a large model encourages you to rely on their large model. But it's worth reading [Anthropic's guide on folding subagents into the main loop](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop).

I dont understand how a single main agent and ever increasing context window can be better than doing granular implementation tasks in a subagent with fresh context.

The real benefit of a subagent is fresh context. The guide reserves subagents for measured context bloat and disposable parallel jobs. In all other cases, a clean slate is paid for with a handoff.

The moment it rewords intent or leaves out an important detail, the subagent has absolutely no idea about this (how would it?) and does the (ever so slightly) wrong thing.

That cost is exactly why my own division is between sessions. [I have one lead session that maintains the ledger and checks the work](https://michael.zazon.us/notes/running-six-to-nine-claude-code-sessions-at-once). Every project has its own peer session, so all intent lives in a single place.

And review has to come from a different model in a different harness, no matter how lean the rest gets.

Having different models from different providers review and collaborate with your main agent gives massively better outcomes. It is slower but the trade off is better and more researched results

I'd fold the helpers into skills and keep the review as a separate model with its own context. It's in the same spirit as the review gates I have in [my agent fleet](https://michael.zazon.us/notes/running-an-agent-fleet-like-a-small-team). Next up I want to try Anthropic's [single-agent loop migration kit](https://github.com/anthropics/single-agent-loop-kit), starting with the evals as the guide suggests.