Member-only story
OpenAI Astra Is Eating Your Agent Harness
The model needs less scaffolding. Your controls need more substance.
On September 3, OpenAI released Astra. A few days later, Eric Provencher published a post on OpenAI’s developer blog called “Rethinking skills and prompts for GPT-6 Astra.”
At first glance, it reads like a list of things to remove from your codebase.
Provencher’s examples are familiar to anyone who has built an agent workflow: a full repository map before every edit, a long list of documents the model must read, or a prompt reminding it to run tests. Those instructions made sense when models needed constant steering. Astra can often work out what it needs for itself.
That creates an uncomfortable possibility. The instructions you added to compensate for a weaker model may now be slowing down a stronger one.
So yes, part of your harness has an expiry date. But that is only half the story.
What is acutally in a harness ?
The model is eating the scaffolding. It is not eating the control plane.
By “harness,” I mean the system around the model: session management, task orchestration, prompts and policies, tool routing…
