Working Surfacelink · 9 Nov 2026
04
A horse in a leather harness pulls a cart along a fenced lane whose gates open only at marked posts, while a figure seen from behind holds the reins and raises a flag at the next gate, the whole scene drawn with wide empty margins on every side.
Nielsen Norman Group

Agent Harness: Definition

Caleb Sponheim · 6 November 2026

How-to · agents · handoff · Agent ownership

Read the original
Takeaway

Caleb Sponheim of Nielsen Norman Group defines the agent harness, the software that sets an AI model's context, tools and limits, and asks designers to specify it alongside the screen.

Summary

Caleb Sponheim, writing for Nielsen Norman Group, which researches and teaches user experience design, defines the agent harness for designers. Two AI scheduling assistants built on the same AI model can differ widely: one only edits meeting agendas, while the other checks calendars and sends invitations. Sponheim explains that the difference lies in the harness, the software that runs the model's requests to its tools and controls what it sees and may do.

Key points
  • The harness runs a loop: the model asks for a tool, the harness checks permissions, calls the tool and returns the result, until the model answers the user.
  • It assembles what the model sees, such as saved preferences. A preference the harness never passes on is lost, and the user has to repeat it.
  • It decides which tools the model may use, when the user must approve an action, and limits on time, cost and steps before the model returns to the user.
  • Sponheim asks designers to decide with engineers which actions need approval and what a cancel stops, and to test the whole task, including a disconnected calendar or a request with no good answer.
Implication

A design team working on an AI feature can treat permissions, approval points and stopping rules as part of the design rather than engineering detail. This is a definition with advice; it reports no team's results.

Plays
Design

For each action an AI agent can take, decide with engineers whether it needs the user's approval and what stops when the user cancels.

Design

In evaluations of an AI feature, test the whole task end to end, not only the text it writes, and include cases where a connected tool is disconnected, a saved preference changes or no answer fits.

Derived by Working Surface from the article; more in the Playbook. Source line: Designing an agent harness involves many human-centered design decisions.

Source issue

9 November 2026: as AI agents remove the slow steps that once gave teams time to question the work, teams are writing down the owners, measures, rules and limits those steps used to supply.