Skip to main content
One agent is rarely enough for a whole job, so you write the layer that runs several. The inner harness is Claude Code, Codex or their kin: one model, one loop. The outer layer runs several of those as a team: it gives each a role, hands work between them, sends work back when a review fails, asks a person at the gate, and keeps a record of every step. Obversa is that outer layer. The team is a file:

What the outer layer owns

Things that catch people out

  • The outer layer is not another agent. It is a process, written down. Nothing in it thinks; the engines do the thinking, one bounded call at a time.
  • A second provider is the point. A review by the same family that wrote the work agrees with it more often. The outer layer is where a seat from another provider is declared.
  • The record is the product’s memory, not the model’s. Every step is a line in a file, readable after the run. See how a run is recorded.

Where to go

What is an agent harness for the inner layer; domain-specific harnesses for a whole job written this way.