Skip to content

Blog

Skills hold what agents need to know, sub-agents do the thinking

The first architecture decision in an agent setup is sorting what needs to think from what just needs to know. Skills carry static knowledge, sub-agents do dynamic work, and the split decides the cost.

Written by

Coding agent setup

Published: 25 June 2026 Updated: 2 min read

An enterprise team showed me their agent setup, and everything in it was a sub-agent. A sub-agent that knew their testing conventions, a sub-agent that knew their architecture criteria, a sub-agent for almost everything. It worked. It was also slow and expensive, and the team could not quite say why.

The missing distinction was simple: the setup never separated what needs to think from what just needs to know.

A skill is knowledge that gets read

A skill is static knowledge, loaded on demand: how this team writes a good test, the criteria a clean architecture has to meet, the release checklist. It loads when the task needs it, does its job, and gets cleaned out of context afterwards. It is cheap and fast, because nothing reasons. It just gets read.

A sub-agent is work that has to think

A sub-agent is dynamic work: researching a pattern across the codebase, analyzing how a module actually behaves, weighing two possible designs. It needs its own context, explores for a while, and comes back with an answer you couldn’t have written down in advance.

How the team sorted its setup

So the team sorted the pieces again. Testing conventions and architecture criteria became skills. Pattern research and codebase analysis stayed sub-agents. One piece earned its place as a sub-agent for a different reason: a refactoring advisor, kept separate so it could review with a clean second pair of eyes instead of inheriting the bias of the work it was checking.

Model choice follows the same question

Most tasks run fine on the small, fast model, and the large one stays reserved for real planning and review. Roughly nine in ten tasks run on the cheap model and the rest on the expensive one, sorted by the same question.

Does this need to think, or does it just need to know? Answer that per piece, and the cost stops being a mystery.

Walk your own agent setup once. How much of it is thinking when it only needs to know?

Further reading

  1. An agent is a loop, and the loop is what you debug A coding agent perceives, decides, acts and observes, then repeats. Naming which of the four steps failed turns a mystery into a fix. 2 min readShort post
  2. A harness is a list of things the model cannot do yet A harness is not plumbing around a model. Every component in it encodes one assumption: a thing the model cannot be trusted to do alone. Read it that way and the harness becomes a checklist. 2 min readShort post
  3. Agent teams win by moving complexity out of the agent Vercel, Manus and Pi got better results by removing tools from the agent. The orchestration belongs in deterministic factory steps around a small, focused agent. 1 min readShort post

All 14 posts on Coding agent setup

ClientsClient work

Scaling needs named owners

Coding agents scale across teams when named people own the platform, the team practice, the guardrails, and the measurement. Rollout & scale is where that ownership is built and handed to internal owners, in your stack.

Training on this topic: Agentic Engineering for Developers and QA
See Rollout & scale

Want to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check