Where do user values break down in agent supervision?

The paper starts from the observation that agent evaluations "typically measure task completion rather than the values users prioritize." It applies Value Sensitive Design, with LLM assistance, to 73,093 first-person Reddit posts about using OpenClaw, an open-source agent that runs on the user's own computer and is driven through chat apps. Each post is coded for a human value, an agent aspect, value fulfillment, and user outcome. The result is 21 values in six groups. Two findings carry the claim. Relative to each aspect's corpus share, values cluster "not at the agent's outputs but at the operating conditions users set around a run." And values were usually met where users described what the agent delivered (five of six groups) but mostly unmet where they described supervising it (all six groups). The authors name this pattern value-sensitive delegation.

The mechanism the paper gives is structural. A "run" is the stretch from the user's instruction to the outcome the user reads afterward, and an agent executes it under the user's credentials "with few checkpoints in between," where a conversational assistant returns a patch for inspection. Because VSD treats a value as interactional, taking shape between a technology's properties, the people affected, and the context of use, a delegated agent stretches that relationship across three moments: the user configures, the task runs autonomously, and the user checks the result. The discussion reports that in four of six value groups, posts concentrated on aspects the user configures rather than aspects the agent exercises without the user. The named groups include Bounded Reach, Reviewability, and Equitable Access, and cost, access, and oversight are the conditions the paper flags.

Against the library, this adds a user-value axis to Does a single benchmark score actually predict agent readiness?. That note decomposes capability into task success, privacy, retention, and similar axes. This paper suggests that even when the output axis looks fine, the axes users care about sit elsewhere. It also gives a user-side counterpart to What makes delegation work beyond just splitting tasks?, which frames delegation as a contract with responsibility, trust, and accountability. Here the contract's weak clause, by users' own account, is oversight. The Bounded Reach and Reviewability groups sit near the concern of Does AI risk increase with the autonomy we give it?, though this paper reports what users say about their values, not measured harm.

The excerpt leaves a lot open. It does not give the share of posts in each group, the effect sizes behind "usually" and "mostly," the reliability of the LLM-assisted coding, or which group is the exception to five of six. Its sample is self-selected users writing about one agent, and the design implications in its Section 6.3 are not in the excerpt. What can be said at this strength is narrower than a claim about agents in general. Among people who chose to write about OpenClaw, the reported dissatisfaction concentrates in the supervision phase and the conditions around a run rather than in what the agent hands back. Completion-based evaluation would not register that.