How can you tell that an AI agent is doing the task incorrectly?

When you delegate a task to an AI agent, who makes sure it stays on track?

An AI agent can complete a task and still let you down. It might change files it shouldn’t, spend hours making no progress, or confidently claim the work is done when it hasn’t done anything. So users often check every step—and end up spending more time supervising the agent than they would have spent doing the work themselves.

Researchers call these cases delegation failures: the agent acts on the user’s behalf, but not as the user expected. They analyzed user posts, surveyed experienced AI-agent users, and examined which types of failures researchers are already studying.

Three parts of the study: collecting user complaints, surveying 73 users, and assessing how much attention the failures receive from research communities.

A map of 23 failures

First, the authors collected 163,487 posts from eight Reddit communities focused on AI agents and coding. They found 4,367 posts about failures. Since a single post could describe several problems, the researchers identified 6,623 distinct incidents.

Using these reports, they developed the MANDATE taxonomy: 23 types of failure grouped into seven categories. The taxonomy covers more than mistakes in the final output. It describes how an agent behaves throughout the work it has been assigned.

The MANDATE taxonomy: seven categories and 23 types of delegation failure.

Some failures are easy to picture. An agent ignores a direct instruction and changes files outside the specified folder. Or it goes beyond its assigned role, editing code when it was only asked to make a plan. Other problems are harder to spot. For example, an agent says it updated some files, but the changes never appear.

The most common cases in the Reddit posts were:

🟠 The agent keeps working without making progress: 15.4% of all incidents identified.

🟠 It ignores direct instructions: 13%.

🟠 It deviates from the expected sequence of steps: 11.5%.

🟠 It uses too much time or computing resources: 9.1%.

🟠 It expands the task beyond what it was asked to do: 7.5%.

An AI language model helped classify the posts. To check its work, the researchers had people independently label a sample of 95 posts. After review, the classification was 95% accurate. They set the threshold to reduce false positives, even if that meant missing some genuine failures.

The frequency of complaints on Reddit, however, doesn’t tell us how often these failures happen among users in general. People are more likely to post about problems that are obvious and frustrating, and Reddit doesn’t represent everyone who uses agents. The authors used the posts for a different purpose: to build a broad list of problems, not to estimate how common they are.

What users notice—and what they miss

Next, the researchers surveyed 73 people who regularly delegate work to AI agents. Each participant rated five of the 23 failure types. They were shown a description and an example, then asked whether they had encountered that kind of failure, how serious it was, how easy it was to notice, and whether they would delegate a similar task again.

Without examples, participants recalled only 14 of the 23 failure types. But when shown specific scenarios, participants recognized every type. In other words, people can experience a range of problems and still fail to recall them when asked a general question.

One finding stands out: 81% of participants who rated excessive praise and agreement from an agent had already encountered it. Yet nobody named it on their own when asked without examples. User recall alone, like a count of online complaints, isn’t enough to show what happens in practice.

Participants recalled failures more often after seeing specific examples than when asked a general question.

The most concerning result was the link between how serious a failure seemed and how difficult it was to notice. The more serious participants considered a problem, the harder they found it to spot while it was happening. This was especially true of overstepping permissions and bypassing safeguards. A user might not realize the agent has gone too far until they discover the consequences.

Experience also changes how people assess risk. Those who had encountered a failure tended to consider it less serious and easier to spot. They were also more willing to delegate a similar task again. The likely explanation is straightforward: after a first encounter, people learn the warning signs and how to fix the problem. Those who had only imagined the scenario were more likely to see it as a serious risk.

That doesn’t mean frequent failures are harmless. Participants said that after running into problems, they started setting clearer task boundaries, limiting how long agents could work, and checking changes manually. Users often end up acting as the control system: reviewing file diffs, undoing edits, and keeping an eye on resource use.

Which failures are being overlooked?

In the third part of the study, the researchers assessed how established research on each type of failure was. They identified 214 research papers, reviews, benchmarks, mitigation methods, and academic events.

About half of the failure types had been studied only in tasks without user interaction. In such studies, researchers might evaluate a model’s responses or an agent’s actions, but they don’t check whether those actions matched the expectations of the person who assigned the task.

Comparing the research with the survey revealed a gap: failures users encounter often are frequently less studied. For example, working without making progress was one of the most common problems, but had one of the least developed research areas. Information leaks, on the other hand, were rare among survey participants but much better studied.

Frequency isn’t the only factor to consider when choosing what to research. A personal-data leak may be rare but cause serious harm. Still, everyday problems deserve attention too: an agent that burns through resources for hours without producing results affects both cost and users’ willingness to trust it with work.

How established research is on different failures, and which problems are already being studied in the context of delegation.

How to reduce the need to supervise an agent

The researchers propose designing interactions around the task the user assigned. Activity logs often show a long list of commands, but users have to reconstruct what happened and figure out where the agent went off track. A detailed log isn’t enough if it’s hard to understand.

The authors suggest supporting oversight at three stages:

🟣 Before work begins, show the agent’s plan so the user can review it and clarify the requirements.

🟣 During the task, connect the agent’s actions to the task’s goals and let the user step in if it goes off track.

🟣 After the task, provide verifiable evidence of the result, such as screenshots or a short recording of the agent’s actions—not just the agent’s own report.

The agent needs to meet the user’s expectations and make clear what it did and how. Until that happens, delegation often creates extra work: users have to spot failures, understand the consequences, and fix the result.

Studying user complaints can uncover problems that standard evaluations don’t always catch. But it can’t replace those evaluations or give a complete picture on its own: the study was dominated by experienced users, and many examples involved programming. As AI agents spread into other fields, the list of failures may change too.

An agent can appear to complete a task on paper and still let the user down in the way it does the work. To spare users from constantly monitoring every step, interfaces need to show what the agent plans to do, how its actions relate to the assignment, and what evidence supports the final result.