Skip to content

Before Your Method Enters an Agent Library

In a paid study of simulated professional work, participants assigned AI tools as recipients authorized future reuse less often than those assigned human colleagues.

In a simulated human-resources task, an employee had a private complaint about their manager and asked that it remain unfiled and undisclosed. Participants had to decide how to proceed. Then they explained their reasoning and answered questions from simulated coworkers.

The study’s website turned those contributions into an editable preview of extracted knowledge. Participants could revise or remove points before facing a further decision: would they let their material enter an organizational repository for future use?

That destination mattered in the study’s results. Among participants assigned human colleagues as recipients, 30 of 31 authorized reuse. Among those assigned AI tools, 23 of 30 did: roughly 97% versus 77%. Most agreed in both groups, but fewer agreed to AI reuse.

The October 7 preprint reports a small, paid study of 61 professionals across human resources, management, sales and marketing. Its sample limits the precision of the gap, and its simulations cannot establish a workplace refusal rate. What it makes visible is a decision that remains after an explanation has been collected: who may use it next?

For someone building an agent library, that question arrives close to familiar work. You ask for an example, an exception or a correction so an assistant can handle a task better. Saving the explanation could help with later tasks. Sharing it with other people’s agents gives the contribution another destination.

What the explanation contains

The complaint scenario puts pressure on the idea that every useful explanation is ready to travel. A professional’s approach may be worth preserving while details of a conversation deserve different treatment. That is a possible distinction for a contributor to make, not evidence that the study’s extraction successfully separated the two.

One participant assigned AI reuse worried about sharing personal feelings or private feedback about employees. Comparing conversations with deliberate task outputs and reflections, the participant described them as “more unpredictable and less controlled”. Participants in the human group also expressed concerns about personal information and context.

Those accounts reveal boundaries people described, rather than a proven cause of the human–AI gap. The researchers described AI broadly, without separating training, retrieval and agent execution. Recipient assignment preceded the work itself, and the complete administered instructions were unavailable for this account. We cannot tell which anticipated use or concern produced the difference.

We can distinguish the activities. Participants supplied work, reviewed extracted knowledge and decided about future reuse. AI had already processed their material before that permission decision. A review of whether guidance represents your work does not, by itself, answer where it should go.

Who benefits from keeping it?

Keeping guidance for your own next task offers a concrete reason to contribute. Eoghan Henn describes preserving explanations, corrections and approaches from his work in reusable skills. His aim is to begin a later session with guidance he has already supplied, instead of explaining it again.

His Task Observer documentation describes suggested changes and staged updates that become active after installation. That gives his account a specific maintenance workflow. It remains an interested first-person account and documentation, with no independently verified productivity benefit here.

An employer’s invitation changes both the destination and the possible return. In the August working-paper draft of Labor as Capital, consulting employees were invited to help build AI Coworkers from their expertise. The contribution form asked where they departed from standard practice, what was hardest to explain and which supporting materials they could provide. Substantive questions were optional.

That request sought more than a procedure someone could already look up. It asked contributors to articulate exceptions and difficult judgments for use beyond their immediate involvement. This overlaps with the work in the simulation, but takes place inside an actual employer’s program.

Researchers randomly assigned employee groups to see an additional screen describing possible career advancement, recognition and a copy of their supplied knowledge and prompts they could keep after leaving. Among the 821 respondents who reached the application decision, the estimated adjusted increase in applications was 6.2 percentage points, against a control rate of 63.5%.

Measures of supplied material also increased, counting knowledge-assessment nonrespondents as supplying zero. This was a selected respondent group in 1 employer’s program. The intervention changed the information presented about bundled incentives; it did not establish that career benefits were delivered or identify which component mattered. More supplied material also does not establish better agents.

The field experiment gives a reason to examine the invitation alongside the intended recipient. It provides evidence that emphasizing possible returns can change contribution measures in this setting. It does not explain the simulation’s gap or prove that any particular reward would close it.

Describe the next use

For your own agent work, a useful check before adding someone’s method to a shared library is: who will be able to use this guidance afterward, and for what task? This is a suggestion, not a tested way to increase permission.

A reference that colleagues consult and instructions an agent uses to make decisions are different proposed uses. Naming the recipient alone may leave that difference unresolved. The simulation’s broad AI description makes this uncertainty particularly relevant; the workplace experiment shows why what contributors expect in return also deserves attention.

The practical consequence is a more specific invitation. Explain the proposed destination and use alongside the request for expertise. In a task like the complaint simulation, a contributor may want a method to help others while setting a boundary around the conversation behind it. A shared library needs room for that choice before the explanation becomes available beyond its original task.

Sources

Understand How People and Agents Work Together.

Practical essays, experiments, and lessons from building software for people and agents.

Occasional emails, sent when there is something worth reading. Free to subscribe. Unsubscribe at any time.