Anyone who has been sent a long pull request or a ten-page document that nobody seems to have read before sending will recognise the problem this manifesto is about. Speakers and attendees of the AIE Code Summit 2026 have published a No AI Slop manifesto, deliberately modelled on the Agile Manifesto that 17 engineers wrote at a ski lodge in Snowbird, Utah, in 2001. It is not a product or a rule, but it gives teams shared words for a cost that coding agents have made very visible.
The format copies Agile closely. There are five value pairs: outcomes over output, understanding over speed, your own thinking over a model's first draft, better work over easier work, and improving the system over patching the output. The definition at its centre is practical: slop is when someone puts less effort into producing something than they ask the reader to spend understanding it. Using AI does not by itself make work slop; passing output along without taking responsibility for it does.
The strongest argument is about asymmetry. The cost of producing a thousand-line change or a long document is falling towards zero, while the cost of reviewing it stays the same, and the authors note that a swarm of agents can produce more in a day than one person could review in a lifetime. They also warn that slop compounds, because the next agent reads and copies whatever sits in the codebase, the wiki and the chat history.
The resolutions are modest and usable. Say when an agent wrote something you are sending, and check the parts that matter. Tell colleagues, kindly, when they have sent you slop. Let linters, type checkers and automated tests catch whatever they can, and fix the system when the same mistake appears twice. Measure what shipped and who it helps, not lines of code, merged pull requests, tokens or agents running in parallel. For teams using Claude Code, Cursor or Codex, that last point is a direct challenge to the usage dashboards those tools show.
The limits are those of any manifesto. It has no enforcement, the signatories are mostly people who already care, and it admits that none of them has this figured out. It also concedes that one-shot prompts and vibe coding are useful. Whether it changes behaviour depends on whether managers, not just engineers, sign up.