Claude chief of staff agent: an AI second in command for a biotech founder
An agent is one job handed to a separate process with its own context, which finishes and reports back. The interesting one is not the agent that does the work. It is the agent that checks it.
The three words in this cluster get used interchangeably and they are not the same thing. A prompt is text you paste. A skill is a packaged procedure that fires when the request matches it. An agent is a delegation: a job, a fresh context window, a defined output, and a hand-back. Anthropic describes both of its local surfaces as able to spin up sub-agents and sustain long tasks, so this is a property of the platform rather than a trick.
What that buys you is real but narrow. A sub-agent can go and read forty files without stuffing all forty into the conversation you are having. It can run a long sweep without derailing what you are in the middle of. And, most usefully, it can look at something with no memory of having produced it.
What an agent adds that a skill does not
| Skill | Agent | |
|---|---|---|
| Shape | A procedure that runs in your session | A job handed to a separate process |
| Context | Shares everything you have said | Starts from only what it was handed |
| Good for | The repeatable step you want done identically | Wide reading, long sweeps, independent checking |
| Cost | Cheap | Time and tokens, so use it deliberately |
That second row is not a technicality. Shared context is exactly why a system asked to check its own work tends to agree with itself: it remembers what it meant, so it reads its own summary charitably. An agent that receives only the claim and the source has nothing to be charitable with.
The one that earns its keep
The risk in handing real work to a machine is not that it cannot do the work. It is that it hands back a confident answer built on something that quietly went out of date, and the person who reads it acts on it. Checking every answer defeats the point of delegating. So the check has to be automatic, and it has to be done by something that did not write the answer.
That is what the verification agent is for. Before an answer reaches you, a second pass re-reads it against the sources it cited and tries to find the place where the conclusion goes further than the evidence. Anything that overclaims comes back downgraded, marked as worth flagging rather than stated as settled. What it catches is usually not a wrong number. It is a correct set of numbers with a wrong inference sitting on top of them.
The shape of the error it is built for. A first pass compares an interim result against a baseline and reports a clean improvement. The checker re-reads the cited files and finds that the baseline was an estimate rather than a measurement, and that the confirming analysis is still open. Every number in the first answer was right. The claim built on them was not, and it was the claim you were about to repeat in a partner meeting.
Why this matters more in drug development than elsewhere
In most businesses a confidently wrong summary costs an awkward correction. In drug development the summaries travel, and they travel into rooms where they turn into commitments. A program status that says the IND package is on track becomes a date said out loud to a board. A read of a partner's obligation becomes a position taken in a joint committee. A view on whether a study met its criteria becomes a paragraph in a diligence response that the other side's team will read line by line and check against the underlying report.
The asymmetry is what makes the checker worth its cost. Being slightly late with an answer is recoverable. Being confidently early with a wrong one, in front of a partner who then finds the gap themselves, is not, and it is the single most expensive thing that can happen in a diligence process. So the design choice is to make the system say "I could not confirm this" often enough to be believed the rest of the time.
When not to use one
Agents are not free. They take time and tokens, so a setup that spawns one for every small task is slower and more expensive than one that does not. The sensible default is the main thread for ordinary work, and a delegated process for two cases only: reading wide, and checking claims that are about to land somewhere durable.
And an agent does not change the boundary. It still drafts rather than sends, it still runs for one person with one mail identity, and it still does not do discovery science. On that last point, since the word "agent" carries a lot of freight in this industry, the honest position on what these do not touch is AI agents for drug discovery, and where they stop.
The other two members of this set are the Claude chief of staff prompt, which is the standing instruction everything reads, and the Claude chief of staff skill, which is the packaged procedure. The platform decision underneath them is Claude as an AI chief of staff, set up in a biotech, the daily-use view from the top seat is an AI agent for biotech CEOs, and the concept they all serve is what an AI company brain is.
A second pass that never saw the first
A conclusion goes to a checker holding nothing but the claim and the files it cited, and the inference that ran past its evidence comes back downgraded rather than stated as settled. That check is hard to describe and plain to watch, so it is what the demo shows.
See the demo