We Built a Discharge Agent. Its Best Feature Is That It Can't Chat.

A ten-step AI workflow where each step works ninety percent of the time does not work ninety percent of the time. It works about thirty-five percent of the time. The failures compound. Andrej Karpathy calls the path to fixing this the "march of nines": getting from ninety percent to ninety-nine costs real engineering, and getting from ninety-nine to ninety-nine point nine costs about as much again. Every nine is the same amount of work as the last one.

That math is why most agent demos look incredible and most agent products quietly disappoint. The demo is the first nine. The business needs the fourth.

For a while the proposed answer to this was "agent skills": write the procedure into a markdown file, hand it to the model, and hope it follows the steps. Skills help. A recent benchmark (SkillsBench) found that curated skills lifted average task pass rates from roughly thirty-four percent to about fifty. That is a real gain, and it is also nowhere near what you need to run an unattended clinical workflow. A skill is a prompt you hope the model follows. You are still hoping.

The thing that actually moves reliability is the layer underneath the prompt: the harness. The harness is the ordinary, non-AI code wrapped around the model. It decides what the model reads, what tools it can touch, what it is allowed to change, and what proof it has to bring back before anything happens. Stripe's internal coding agent is the clean example. It does not just write code and hand it over. The harness runs the changes against the test suite in a sandbox and makes the model iterate until they pass. That is how they merge over a thousand AI-authored pull requests a week. The reliability is not in the model being clever. It is in the rails refusing to let a bad result through.

We build the same way at WOO, in a setting where the stakes are higher than a failing test. Here is one we are building now.

The discharge agent

When a patient leaves the hospital, they go home with a discharge note: a dense, jargon-filled document describing their medications, follow-ups, and warning signs. Misunderstanding it is one of the reasons nearly one in five Medicare patients is readmitted within thirty days. So the workflow is high value and genuinely messy, which is exactly the profile that tempts people to throw a chatbot at it.

We did the opposite. A caregiver uploads the discharge note. The agent returns a plain-language summary, a structured medication list, and a set of follow-up tasks. The caregiver reviews each item, sets it up, and assigns it. That is the whole product. Three design decisions are doing the heavy lifting, and none of them is the model.

The first feature is the one we removed: there is no chat. No open text box, no "ask the agent anything." A conversational surface is the most flexible thing you can give an AI and the hardest thing to keep reliable, because flexibility is just another word for the number of ways it can go wrong. We took it away on purpose. The agent has exactly one job with exactly one shape of output. That single subtraction eliminates an entire category of failure before it can happen.

The second: every item has to be quoted. Each medication and task the agent surfaces carries a verbatim "From your note" line pulled straight from the source document. "Take amoxicillin 500 mg by mouth twice daily for 7 days." The agent cannot present a suggestion it cannot quote. Provenance is not a nice-to-have feature bolted on top. It is a rail. An ungrounded claim has nowhere to render, so it does not get made.

The third, and the one I am proudest of: the agent's confidence controls what it is allowed to do. Extractions come back in three tiers. High and medium confidence items are selectable and can become reminders. Things the agent spotted but is not sure how to handle (a cardiology appointment, wound care) get surfaced for manual handling but are never auto-created. And anything it could not confirm is marked display only and physically cannot become a reminder. The agent's own uncertainty is wired directly to its permissions. A low-confidence guess does not get an action. It gets a footnote.

Around all of that sits the human gate. The agent proposes a structured list; a caregiver reviews it, unchecks anything they do not want, and confirms each item before a single reminder is scheduled. Nothing applies in bulk. Schedules are edited in exactly one place. When the agent notices a possible drug interaction it flags it in plain language and still leaves the decision to the person.

Reliability lives in the box, not the brain

Notice how little of this depends on the model being smart. The model does two things: it rewrites the note in plain language and it extracts structured items. Everything that makes the system trustworthy is the box around it. The schema that data has to conform to before it lands in the vault. The audit path every write is routed through whether the model remembers to or not. The confidence tiers. The human confirm. The agent is a free-handed drill. The harness is the jig that makes every hole land in the same place anyway. You do not get reliability by upgrading the drill. You get it by constraining where the bit can go.

The honest tradeoff: a harness buys reliability by spending flexibility. Our discharge agent cannot improvise, cannot field a question it was not designed for, cannot handle a document type we did not plan for. In a consumer toy that would be a weakness. In a clinical workflow where a confident wrong answer can send someone back to the hospital, the inability to improvise is the entire point.

Right now a caregiver does the assigning. The next step is widening the agent's reach, eventually to clinical staff and more automation. That widening is not a prompt change. It is a harness change, and every bit of reach you hand over has to be earned by reliability you can prove. That is the harder, less glamorous half of this work, and it is the next thing I want to write about.

If you are building agents for a regulated, high-stakes workflow, the question is not whether you can build one. You can. The question is whether you can build the box around it. What is the first thing yours should not be allowed to do?

Building AI infrastructure for regulated industries? Get in touch.

So What is Next?