The harness decides more than the model

In the same week in July 2026, OpenAI launched ChatGPT Work, an agent built on Codex and GPT-5.6 Sol that can act across a user’s connected apps and files, stay on a project for hours, and turn a stated goal into finished documents, and The Information reported that Cursor has been running an internal general-purpose office agent, codenamed Sand, since late June, with a public release still unconfirmed. Neither story was really about which model is smartest. Both were about the same underlying shift: coding tools spent two years building the scaffolding that turns a capable model into finished, trustworthy work, and that scaffolding is now the product being sold to everyone else.

What a harness actually is

A model on its own answers questions. A harness is what turns that into work getting done: the tools it is allowed to call, the context it is given before it starts, whether the task runs as one agent or several narrower ones, and what happens when it reaches the edge of what it can determine. ChatGPT Work makes this concrete. OpenAI recommends learning it on familiar tasks, defining a goal, supplying the relevant context, then supervising a large multistep task by answering questions or changing direction rather than checking every step. It connects to Notion, Google Drive and Microsoft 365 so the agent can reach real work context instead of whatever a user happened to paste in, and it supports scheduled cloud tasks that keep running after a laptop closes. None of that is a claim about GPT-5.6 Sol’s raw capability. It is a claim about the environment built around it.

Early users are already grading the harness, not the model

The most interesting reactions to ChatGPT Work were arguments about the harness, not the model underneath it. Dan Shipper called the merged Work and Codex desktop app “the gold-standard for knowledge work in AI” largely because of how the interface handles a long-running task, not because of a benchmark score. Peter Yang criticised the opposite problem: the split between Chat, Work and Codex, and the inconsistency across web, desktop and mobile, calling it messy enough to get in the way of the underlying capability. Ethan Mollick said the trade-off between Work and Codex was unclear to him, and pointed at a different harness question entirely: that using these tools well now means thinking like a manager assigning work to agents, rather than a user chatting with one. Three informed observers, three different verdicts, and every one of them about the scaffolding around the model rather than the model itself.

The pattern was not built for chat

This discipline has a lineage. Coding agents had to solve tool access, context supply and escalation first, because an agent with the wrong permissions in a real codebase does visible, immediate damage, and one with too little context produces code nobody can ship. OpenAI’s ChatGPT Sites feature, which lets an agent turn a piece of knowledge work directly into a hosted web app shared by URL, is the same discipline pointed at a new output: the harness now includes not just what the agent can read, but what it can publish, to whom, and under what access controls. None of this is unique to OpenAI. It is the shape every serious agent for general work now has to take, whichever vendor builds it.

What this changes for anyone choosing a tool

A benchmark chart answers “how capable is the model”. It does not answer “what can this agent actually reach on my systems, on what basis, and what happens when it is wrong”, and that second question is usually the one that decides whether a deployment is safe to trust with real work. An organisation evaluating an agentic product for the first time gets more useful signal from asking what tools it connects to and how narrowly, what context it starts each task with, whether a long task splits into scoped sub-steps or runs as one unsupervised block, and what it does when it is uncertain, than from any single capability score. The model is one input to that picture. Increasingly, it is not the deciding one.