Every: How GPT-5.6 changes knowledge work

Dan Shipper’s July 2026 piece in Every’s Chain of Thought newsletter does something unusual for writing about AI tools: it resists the task-level framing entirely. Rather than asking what GPT-5.6 can do — summarize a document, draft an outline, answer a question — Shipper asks what it makes possible at the system level, for people who do sustained knowledge work over time.

The argument is that GPT-5.6 is fast and affordable enough to build what he calls loops: ongoing systems that continuously gather relevant information, surface decisions to the human who owns them, and execute approved actions while accumulating context from prior rounds. The job of the person running the loop is not to do each task directly but to tend the system that does it — reviewing what came in, approving or redirecting what goes out, and feeding back what worked so the system improves.

Shipper runs through the domains where he and his team at Every are applying this: hiring pipelines where the model screens applicants and prepares briefings before a human reviews them, editorial planning where the system surfaces topic candidates based on reader behavior and asks for direction, customer research where interviews are processed and clustered before a human synthesizes findings. In each case the human is present and deciding, but the gap between raw input and decision-ready output is closed by the loop rather than by manual processing.

For writers and content teams, the piece is useful in a few specific ways. Shipper’s description of editorial loop-building — where an AI system continuously pulls from reader signals, previous coverage, and search data to propose what to write next — is a credible model for any publication trying to be consistent without being reactive. The key constraint he names is also worth holding onto: loops degrade quickly when the model’s context window starts filling with outdated or poorly structured history, so maintaining them requires periodic cleanup rather than just setup.

The piece also introduces Tend, an open-source tool the Every team built to make loop construction more accessible. It is designed to connect to Gmail, Slack, and other information sources and to accumulate feedback over time so the model’s proposals get better as it learns what a particular person or team actually approves.

The framing Shipper offers — “knowledge economy” shifting to “allocation economy” — is more than a phrase. It describes a change in where the expertise actually sits: less in the act of producing output, more in the judgment calls about what output to keep, modify, or discard. For writers who feel the tools are beginning to outpace their process, this article gives a concrete model for what it looks like to get ahead of that curve rather than behind it.