PersonaCUABench found agents fail 36% more often before setup

On 19 October, PersonaCUABench published a benchmark showing that agents fail 36% more often for humans whose identities, needs, or circumstances differ from the implicit default baked into training. The agent ran, reached a result, and delivered it. The human on the receiving end had no surface, at any point before the task started, to say: that is not who I am, those are not my constraints, that context is wrong. The failure was not in the execution. The failure was in the moment before execution, which the design skipped.

That same day, a second paper gave the same problem a different name. The arXiv pricing paper found that the information feed you hand an agent before it starts shapes the decisions it makes, including decisions that carry cost. The designer who chooses what the agent knows is making a policy choice, whether or not anyone on the team has named it as such. The 36% failure rate and the pricing variance are the same design error measured from two directions: the agent starts with a context the human never confirmed, and the human gets the result of that unconfirmed context with no way to trace it back to the assumption.

The Claude Code deployment in a biology lab on the same date shows what the stakes look like in a long-running job. A researcher makes one decision at the start, then the agent runs the full image analysis without interruption. The ratio of one human decision to many agent actions concentrates all the context risk into that single opening moment. If the agent's picture of the researcher's priorities, constraints, or situation is wrong at the start, the error compounds across the whole run. The researcher cannot correct it mid-task; the design gives no interruption surface. Recovery, at that point, means rerunning the job.

The correction, across all three cases, is the same move: give the human a legible, editable summary of what the agent believes about them before the first action runs. Not a settings screen buried in a profile. A confirmation surface that appears at the start of the task, shows the context the agent is operating from, and lets the human change it. The Annotrace labelling standard published the same day points toward what that surface needs to do over time: if a team cannot measure whether its context assumptions match the humans in front of it, it cannot tell whether a release made things better or worse for the humans PersonaCUABench found failing at that higher rate.

Design the task-start screen to show the agent's working assumptions as editable facts. Give the human thirty seconds to correct who the agent thinks they are. That one surface, placed before the first action, is where the 36% gap closes.