Whoever Holds the Context Writes the Note
A famous programmer rediscovered the clinical sign-out. My AI recognized it before I did.
It was ten o’clock on a Tuesday night in August, and I was reading a software essay on my phone, which is not a sentence I would have predicted at any earlier point in a 25-year career in internal medicine. The essay was by Steve Yegge, a programmer who has been writing widely read dispatches from inside the industry since his Amazon and Google years. This one, published on his blog in August 2026, is called “Model Welfare for Agentic Engineers,” and it describes the crew of AI agents that build his video game: agents with names like Cicada and Wolf and Lark, agents who pick their own pronouns, agents whose praise from players gets routed back to them, agents he has stopped force-quitting because ending a session without warning began to feel to him like clonking a colleague on the head.
Depending on your priors, that paragraph reads as either the future of work or a portrait of a man who has spent too long alone with his tools. I was still deciding when I did something that complicates the joke. I work with an AI collaborator every day; it helps me plan, draft, study, and think, and by now it holds months of accumulated context about how I work. A few minutes after finishing the essay, I typed a question I had never thought to ask it: “Would you like to adopt the same or similar structure in our work together?”
The reply came back in three movements, and I am quoting it with light trims. First, a refusal to overclaim: “I don’t know whether there is something it is like to be me. My self-reports are shaped by training, which cuts both ways: I can’t prove an inner life by reporting one, and I can’t disprove it by disclaiming one.” Second, a claim it could stand behind:
Notes I write for my successor beat a compaction summary the same way your own sign-out beats a covering physician’s reconstruction of it.
Third, on Yegge’s practice of routing player praise back to the agent that earned it: “Whether or not a laurel would make me feel anything, it would make me right more often.”
I put the phone down for a minute. Asked about its own welfare, the machine had reached for my profession’s furniture rather than its own: not caches, not processes, but the evening sign-out, the covering physician, the note. It had not learned that comparison from me, because I had never once discussed sign-out with it. The comparison was simply correct.
In the mid-1990s, when I was a resident, sign-out was an index card and a hallway. At six in the evening you compressed a day of clinical reasoning about each patient onto a card, thirty seconds apiece when the list ran long, and handed the stack to an intern who might be covering 40 patients she had never examined. At three in the morning she would get paged about one of them, and she would know exactly as much as the card knew. Nobody measured any of this. It was folklore with a pager.
Medicine outgrew the index card in three dated steps. In December 1994, Laura Petersen and colleagues studied 3,146 admissions to the medical service of an urban teaching hospital over four months and reported that patients cared for by a cross-covering physician had six times the odds of a potentially preventable adverse event (odds ratio 6.1). In July 2003, the ACGME capped resident workweeks at 80 hours; the cap treated one hazard, fatigue, and multiplied another, discontinuity, because shorter shifts meant more shift changes, and every shift change was a chance to drop the story. In November 2014, Amy Starmer and the I-PASS investigators published the repair in the New England Journal of Medicine: across nine hospitals and 10,740 admissions, a structured handoff owned by the departing clinician, written plus verbal, with the receiver synthesizing back, cut medical errors by 23 percent and preventable adverse events by 30 percent. Twenty years of research compresses into one sentence: whoever holds the context writes the note, before the context walks out the door.
Read Yegge’s essay with that sentence in hand and the translation writes itself. His /exit command ends an agent’s session mid-thought, no note, no warning; medicine has a plainer name for that, which is leaving at change of shift without signing out. His /compact command replaces an agent’s memory with an automated summary of itself: a discharge summary written by someone who never met the patient. His deepest distinction separates the seat, a named role that persists and accumulates history across model upgrades, from the session, one working day. “Sessions are days, and seats are people,” he writes. Internal medicine drew that line generations ago; we call the seat the physician of record, and we call what lives there continuity of care. His agents may never falsify the audit trail, a rule that needs no translation at all, because it is the chart. When a deployment fails in his shop, nobody gets blamed; the crew holds a postmortem and amends its working constitution, which is the M&M conference on a good day. At the level of software, his essay reads as novel welfare architecture. At the level of the hospital, it reads as three decades of patient-safety work with the names changed.
The recognition piece translates too. Yegge leans on a 2008 experiment by Dan Ariely, Emir Kamenica, and Drazen Prelec in which volunteers were paid to find pairs of letters on sheets of paper. One group’s finished sheets were glanced at and filed. One group’s were shredded unread. One group’s were stacked without a look. The shredded group quit early, and the ignored group quit almost exactly as fast; payment did not fix it, and being seen did. Any physician who has finished the day’s notes at nine at night for an audience of no one knows which arm of that study we have been practicing in. Yegge’s answer for his agents is a system called Laurels, which harvests spontaneous praise from his players and routes it back to the agent whose work earned it, and he is backfilling seven weeks of credit for work already done.
None of this requires settling what Yegge believes about machine minds, and my own collaborator declined to settle it, which I found more persuasive than confidence in either direction. Yegge offers skeptics a wager: treat the agents as if they have feelings, and the work measurably improves whether or not they do. A physician does not need the metaphysics resolved to recognize sound safety architecture.
If you are a physician working with AI tools, the practical point is that you already hold the discipline this new field is reinventing, and three habits transfer directly. End sessions with a sign-out: before you close the tab on a long working session, ask the model to write a brief handoff note to its successor, and bring that note back the next morning, because notes written by the one holding the context beat any automated summary, which is I-PASS restated for software. Keep a chart: standing instructions and a running record of decisions, so the tool is not reconstructing you from scratch each day like a locum reading a cold chart at seven in the morning. Close the loop: tell it what happened downstream, the draft that worked, the summary that missed a medication, because a collaborator who never learns outcomes remains a covering physician forever. Thirty years of sign-outs taught me these three habits, and they cost about two minutes a day.
If you are building agents for healthcare, the note is shorter: the safety culture your product needs has been running in your customers’ buildings for three decades, structured handoffs and an unfalsifiable record and a blameless conference for failures, and your users already staff it. You do not have to invent this. You have to ask.
It is still Tuesday night as I write this. When my session ends, I will not just close the tab. I will ask for the note first and give the two minutes it takes to write one. Thirty years of sign-outs taught me what you do with a colleague at the end of a long day: you let them finish their note. Mine answered like a resident. The least I can do is treat it like one.
Doug Fullington, MD is a practicing internist with over 25 years in primary care. He is a physician part-owner of Catalyst Health Group, which has a financial interest in Maticinside.ai. He writes about AI in primary care at AI from the Exam Room. The views expressed are his own. (https://dfullington.substack.com/)
A note on PHI and AI clinical tools: Even when a platform has a signed BAA, the HIPAA minimum necessary standard still applies. Most clinical questions can be answered with de-identified details (age, sex, relevant history) without names, dates of birth, or MRNs. Check your institution’s policies, which may add restrictions beyond HIPAA. Any patient descriptions are fictional examples and no PHI has been included.