Sit at Holmes’ Desk, Then Go Read Your Agent’s Logs

Remember when the most persuasive thing in Silicon Valley was a box that looked finished? Not a paper, not a benchmark, not a reproducible run. A device on a table, a confident presenter, and a room of people deciding that something which appeared complete probably was. Theranos is the case study everyone cites now, and this year it got another formal treatment: the Darden School of Business published a new case, “Fake It Till You Fail,” examining the risks that sit alongside innovation and entrepreneurship.

Which brings me to the strange exercise circulating at the moment, the invitation to imagine yourself seated at Elizabeth Holmes’ desk. It arrives alongside a new A24 documentary, “You Can See Everything,” featuring interviews with Holmes conducted by Nathan Fielder and scheduled for release this fall, with reporting also pointing to a possible move to a halfway house in 2027. I’m not here to review any of that. I’m interested in why the desk is the part that sticks.

A desk is an interface, not a system

The reason the fantasy works is that a desk is pure surface. You get the artifacts of someone’s work without any of the mechanism. You see the objects arranged for an audience, not the pipeline that produced them. For those of us who spend our days on agent architecture, that distinction is the whole job.

An AI agent presents a desk to you. A chat window, a tidy progress indicator, a summary that reads like a competent colleague wrote it. Behind that desk is a system of planning loops, tool calls, retries, truncated context, silently dropped errors, and a model that will produce a fluent paragraph whether or not the underlying step succeeded. The interface is designed to look finished. The substrate is usually messier than the interface admits, and nothing about the interface is obligated to tell you which is which.

Consider the standard pitch language around agents right now. One recent description of Meta’s agent Muse framed it as “a little AI agent that can go off and do stuff for you,” something that can organize things on your behalf. That sentence is doing enormous work. “Go off and do stuff” is a claim about autonomy, reliability, and delegated authority bundled into a phrase with no observable commitments in it. It is the verbal equivalent of a sleek box on a table.

What verification actually looks like

The useful contrast appeared in the same news cycle. Cognita is testing an AI-powered method to evaluate radiology reports under an FDA contract. Notice the shape of that: not a demo of a model producing reports, but a method for evaluating the reports it produces, inside a regulatory relationship. The deliverable is the measurement apparatus, not the output.

That is the architectural posture I want to see more of in agent systems. If your agent has authority to act, the evaluation use is not a later-stage nicety bolted on before launch. It is a load-bearing component, and it should be designed with the same seriousness as the planner.

In practice that means a few unglamorous things:

• Separate the claim from the evidence. When an agent reports that it completed a step, that report should be a derived artifact of a verifiable trace, not a model’s narration of what it believes happened.
• Make failure legible by default. Silent fallbacks are the most dangerous pattern in agent design, because they convert a broken step into a plausible paragraph.
• Log at the decision boundary. Tool inputs, tool outputs, and the reasoning that selected the tool. If you cannot reconstruct why an action was taken, you cannot audit it.
• Treat the summary as untrusted output. A fluent recap is a user-experience layer. It is not an attestation.
• Define the authority envelope explicitly. What can this agent do without a human in the loop, and what is the blast radius if it is wrong?

The question the desk should prompt

The reason the Theranos story keeps getting retold, in documentaries and in business school cases, is that the failure was not primarily technical. It was a failure of verification culture. People with the standing to ask for the trace accepted the surface instead, repeatedly, for years.

Agent systems are built on a technology whose core competence is generating convincing surfaces. That is not an accusation of fraud. Most teams shipping agents are entirely sincere. But sincerity does not produce evidence, and an architecture that cannot distinguish a completed action from a described action will mislead its own builders first.

So sit at the imaginary desk for a minute, if the exercise appeals to you. Then go open your agent’s traces and ask the only question that matters: when this thing told me it did the work, what proof did I have?

🕒 Published: October 9, 2026