Saturday
3 October 2026

Teams that let AI agents build are finding that the instructions and examples those agents read have become a product in their own right. Where someone owns that material and keeps it current, agent work stays consistent. Where nobody does, the setup grows into a tangle that only the AI can find its way through.

Tanner Kohler of Nielsen Norman Group (3 Oct · 01), a user experience research and consulting firm, studied people without engineering backgrounds who build their own AI agent setups for daily work. He found them growing these setups piece by piece, patching failures and dropping files into unsorted folders, until they could no longer say where things lived. One product lead had built a library of agent instructions that his whole team was required to use. He was rebuilding it for the second time because it had outgrown what he could keep in his head. Kohler advises keeping three kinds of material apart: instructions that apply everywhere, material for one task, and streams such as email that the AI can read. He also recommends reading the system-wide instructions from start to finish now and then, so that they stay short enough to review in one sitting.

Luke Wroblewski (3 Oct · 02) is a product designer who has worked on design systems for more than 20 years. He starts from an old problem: the design system ends up describing a product that no longer exists. With AI agents, marketing staff, engineers and product managers can all change a website, and each person's agent will choose its own colours and spacing unless something steers it. On Intent, one of his recent website projects, a design lead and a front-end lead defined the grid, fonts, colours, spacing and components once. An instruction file for agents, called AGENTS.md, tells every agent to reuse those values instead of inventing new ones. The team treats the code as the reference, so the design system page built from that code stays current.

At Stripe, the payments company, engineers on the Local Payment Methods team (3 Oct · 03) faced integrations that could take six months each, across more than 125 payment methods. David Dunne and Xenofon Vourliotis, software engineers, and Sai Samant, a technical writer, describe a "factory" built from more than 100 reusable prompts that cover each step of an integration. In a side-by-side test, one task took about 20 days of engineering effort with a general coding agent and 4 days with the saved prompts. To keep the prompts current, each run collects what the agents had to learn, the decisions they made alone, and the engineers' review comments. An AI agent then proposes prompt changes from them. The team also moved coordination out of an AI agent and into ordinary code. The agent had been slow, costly and sometimes did the work itself instead of handing it on.

Sean Goedecke (3 Oct · 04) is a software engineer who writes essays about engineering practice. He argues that the engineer working beside a coding agent is no longer there to make the code more correct. In his account, agents already write code that compiles and rarely contains concurrency errors. Left alone, however, they produce work that suits the company badly: hard to maintain, and trading away real requirements to satisfy invented ones. He sees his job as aligning the agent with his organisation's technical values, which differ from one company to the next. His practical advice is to state those values to the agent directly.

In each case, the knowledge that keeps agent work consistent lives in written material that a person or team has to own. That material is a context library, a design system, a set of prompts or a statement of values. Stripe and Wroblewski show that material improving because a named team maintains it from real work. The Nielsen Norman Group study shows what happens when nobody does, and Goedecke explains why the material is specific to each company and cannot be bought ready-made.

Stripe and Wroblewski describe their own successes, while the tangled setups in the study belong to individuals working alone with no review at all. The contrast may therefore say more about team size and engineering skill than about ownership of the material.