Where Values Actually Live When Users Delegate to AI Agents

When an AI agent executes a task on your behalf, what matters to you? Technical evaluations answer this with completion rates and benchmark scores. But a task that completes successfully can still exhaust a budget you cannot replenish, access files you never intended to expose, and leave you unable to explain what happened. The values users care about when they delegate are not always the values the agent was optimized to deliver.

This paper, from researchers at the University of Cincinnati, Penn State, University of Arizona, University of South Carolina, and Florida International University, asks what human values surface when real users describe their experience with OpenClaw, an open-source autonomous agent that runs on a user's own computer and performs tasks through chat apps like WhatsApp. Released in November 2025, OpenClaw drew more than 370,000 GitHub stars by August 2026. Its users post about it on Reddit constantly. The authors analyzed 73,093 first-person Reddit posts, each coded for its human value, the agent aspect it attaches to, whether the value was fulfilled, and the user outcome.

The framework is Value Sensitive Design (VSD), which holds that design should account for human values throughout development, defining a value as what a person or group considers important in life. VSD argues that technology more readily supports some values and hinders others, and that whether a value is realized depends on the goals of the people interacting with it. Applied to autonomous agents, this means the question is not just whether the agent completes a task, but what the user experiences while the agent acts on their behalf.

What OpenClaw Users Actually Talk About

The 21 values identified in the posts form six groups. Autonomous, Dependable, and Affordable Operation covers the user's own goals and state while the agent acts, whether the agent's service holds and can be repaired, and what a run costs in money and metered resources. Bounded Reach covers security, privacy, identity, property ownership, data sovereignty, and contextual integrity. Reviewability covers transparency, meaningful human control, accountability, and informed consent. Equitable Access covers universal usability and freedom from bias.

Five of the 21 values account for 70.8% of all posts. Autonomy is the most frequent at 19.9%, followed by dependability at 19.8%, affordability at 14.0%, resource stewardship at 8.9%, and universal usability at 8.3%. Four of the remaining 16 values appear in fewer than 250 posts each.

The most striking finding is where these values attach. The agent aspects most overrepresented relative to their corpus share are not the model core, not the output, not the task completion, but the operating conditions users set around a run: what it cost, what it could reach, what it recorded, when it had to ask, and how it was installed. The model core, the underlying LLM that produces the agent's output, is not among the most overrepresented aspects. Values cluster not at the agent's outputs but at the boundaries users set around delegation.

Values Are Met at Delivery, Not at Supervision

The fulfillment rates tell the story. Values were usually met where users described what the agent delivered, in five of six groups, and mostly unmet where users described supervising it, in all six groups. The post's value was met in 67.7% of task-effectiveness posts but in only 34.7% of resource-burden posts and 10.7% of risk-exposure posts.

The split is sharp. When a user talks about what the agent accomplished, the value is usually met. When a user talks about what it cost, what it accessed, or what they had to watch over, the value is usually not met. The two largest value groups, Autonomous Operation and Dependable Operation, both share one agent aspect, the model core, but their fulfillment patterns diverge. Autonomous Operation exceeds its expected met rate at 77.3%. Dependable Operation, which includes trust, dependability, and repairability, does not.

Resource accounting, the agent aspect tracking what a run consumes and what it is billed for, has the lowest met rate of all 18 agent aspects at 35.3%. This is the aspect users interact with least during a run but worry about most, because cost accrues where they cannot see it.

The Six Value Groups in Practice

Autonomous Operation, where users widen reach rather than take it back. Users report that the agent did what they asked, but the autonomy they value is not the agent's independence. It is their own, the ability to set a task and walk away. The outcome that matters is not completion but what the user can do while the agent works. Users describe reaching capabilities they could not have accessed manually, not because the agent taught them but because it did the work while they were free to do something else.

Dependable Operation, delivery that did not hold across runs. The most common dependability complaint is inconsistency: the same prompt produces different results on different runs, or the agent's output looks correct but is subtly wrong. Users describe spending more time verifying than they saved by delegating. The value of dependability is not met when the user cannot predict whether a run will succeed, even if individual runs do succeed.

Affordable Operation, cost that accrued where users could not see it. Users describe surprise at token consumption, API charges, and metered resource usage. The cost is not visible during a run. It appears after, in a bill or a depleted allowance. Users who describe affordability as met set explicit budget limits before running. Users who describe it as unmet did not, or set limits that turned out to be insufficient. The value is about predictability of cost, not its magnitude.

Bounded Reach, risk and usefulness both settled at setup. Users set file system access, network permissions, and credential scopes before a run. Those settings determine both what the agent can do and what it can break. Users who set tight bounds find the agent less useful but safer. Users who set wide bounds find the agent more useful but harder to recover from when something goes wrong. The value is met when the setup matches the user's actual risk tolerance, which is hard to know in advance.

Reviewability, checkpoints that worked at the run's edges but not inside it. Users describe approving a task at the start and reviewing output at the end, but having no visibility into what happened between. The agent's intermediate steps, tool calls, and file modifications are not surfaced in a way users can inspect. Reviewability is met when the user can explain what the agent did. It is not met when the user can only see the result and must trust that the process was correct.

Equitable Access, useful work reported from the far side of setup. Users who get OpenClaw running describe productive use. Users who cannot get past installation, configuration, or API setup describe nothing. The value is met for those who clear the setup barrier and unmet for those who do not, and the barrier is not the agent's capability but the user's ability to configure it.

Why This Matters for Agent Design

The pattern the authors call value-sensitive delegation reframes where value-alignment work should focus. Current agent evaluation measures task completion and trust calibration. These measure the model core and the task specification. They do not measure resource accounting, environment access, observability, or system access. The values users report as unmet concentrate at these operating conditions, not at the model's output.

The implication is that supporting human values requires attention not only to what an agent accomplishes but to the conditions users set around delegation. A complete task is not a successful delegation if the user cannot afford the next run, cannot explain what happened, or cannot recover from what went wrong. The agent's output is the part that works. The surrounding conditions are the part that does not.

The authors propose design implications keyed to the agent aspects where shortfalls concentrate. Cost visibility needs to be real-time, not post-hoc. Access controls need to be comprehensible at setup, not just enforced at runtime. Reviewability needs intermediate checkpoints, not just start and end. Observability needs to surface what the agent did, not just what it returned.

For developers building autonomous agents, the practical takeaway is that the operating conditions around a run are where users' values live. Optimizing the model core improves task completion. Optimizing the boundaries around delegation improves the user's experience of delegating. The two are not the same thing, and a benchmark that measures only the first misses the second entirely.

The paper also demonstrates a method for studying agent experience at scale. Using LLM assistance to code 73,093 Reddit posts, validated against human coders, the authors produce quantitative patterns and qualitative themes simultaneously. The human validation shows high accuracy (0.900 to 0.994) but variable macro-F1 (0.604 to 0.924) across coding targets, which is a useful calibration for anyone considering similar large-scale coding approaches.

The posts are self-reported, the unit of analysis is the post not the user, and the platform is Reddit, which skews toward technically literate users who can install and configure an open-source agent. These are real limitations. But the pattern is clear enough to redirect attention: the values that matter in everyday agent use are not the values that agent benchmarks measure.

Read the paper on arXiv