Skip to content
Article Nielsen Norman Group Sep 2026

Nielsen Norman Group: three context roles for AI agents

Published in September 2026, Tanner Kohler’s article on Nielsen Norman Group examines how experienced AI users manage information when working with agents. The central observation is that effective agent use depends less on crafting well-worded prompts in the moment and more on how supporting material is organized beforehand. Kohler identifies three distinct roles that context plays across this kind of work.

What the article is about

The article draws on research into how power users structure information when delegating tasks to AI agents. Rather than treating context as a single undifferentiated mass of background material, Kohler distinguishes three types by how they are used and where they belong in the workflow.

The three context roles

Global context covers stable reference material that applies across tasks: brand standards, style guides, governance rules, reference documents. It does not change from one project to the next. Power users build this layer once and maintain it, rather than restating it per session.

Local context is task-specific information: a current project brief, a change log, a set of meeting notes relevant to one deliverable. It is often read and updated by the agent as work progresses. Unlike global context, local context changes as the task moves forward.

Ambient context is the unfiltered information stream that might become relevant: email threads, Slack channels, recorded meeting transcripts. It is not curated in advance. The agent draws from it as needed, reducing the overhead of deciding what belongs in local context before work starts.

Key takeaway

Kohler’s framing shifts the preparation question from “what should I put in my prompt?” to “where does this information belong in my system?” The same piece of information can move between roles over time: a Slack message might start as ambient material and get promoted to local context once someone flags it as directly relevant to the current task.

The article makes a practical argument for separating these information types into distinct locations — not for organizational tidiness but because agents perform better when the category of each input is clear. Mixing governance rules with current project notes makes it harder for an agent to apply one consistently without being distracted by the other.

The shift the article describes is toward curating the right context rather than writing better prompts. A designer who has invested in a well-structured global context — one that captures the design system, the product principles, and the audience — can delegate tasks with shorter, more direct instructions and get more consistent results.

Who it is useful for

Designers, content strategists, and product people who work with AI agents regularly and want to reduce the friction of session setup. The framework applies wherever agents are given substantial context: Notion workspaces, Claude Projects, custom GPT setups, or any other configuration where the agent needs background information to work well. No technical background is required to apply the three-role distinction.