At the end of a shift, one young doctor hands their patients to another. The one going home may be tired and rushed. The one staying will care for those children next, and much of what they need to know will come from the next few minutes.
Christopher Landrigan and Amy Starmer, pediatricians at Boston Children’s Hospital, saw a weak link in that moment. Residents were drilled to take patient histories with templates and to present cases in a set format on rounds. At shift change, when a patient passed to someone new, that structure was largely absent. And rules limiting residents’ working hours had made the moment more common: shorter shifts meant more handoffs.
They decided to give the handoff a structure of its own. The obvious worry was time, a fixed script for every patient at the end of a long shift.
Five letters
The routine they built is called I-PASS. Illness severity comes first. Then a patient summary, an action list, and situational awareness and contingency planning. It closes with synthesis by receiver: the incoming doctor summarizes what they just heard to the doctor who said it.
It came as a package: the mnemonic for spoken and written handoffs, communication training, a training and observation program for faculty, and a campaign to keep it going. The team took it to pediatric residency programs at nine hospitals. Across 10,740 admissions, reported in the New England Journal of Medicine in 2014, medical errors fell 23% and preventable adverse events, errors that actually hurt a child, fell 30%. Harm that better communication couldn’t have prevented didn’t change significantly, which a co-author read as a sign that the handoffs themselves made the difference. Spoken handoffs took 2.4 minutes per patient before and 2.5 after, not a significant difference. It was a before-and-after study of the whole package, and the paper says the program was “associated with” fewer errors.
Handoff time didn’t rise significantly. Alongside the training, what changed was which facts were handed over, in what order, and the check at the end that they had landed.
What the next one knows
An AI model is the incoming doctor on every shift. It knows what it learned in training and whatever text it is handed now: the instructions, the documents, the conversation so far, the results of the tools it has used. Engineers call that text the context. In 2025 Anthropic described “context engineering” as “curating what will go into the limited context window from that constantly evolving universe of possible information.”
More is not automatically better. Researchers at Chroma tested 18 language models on tasks held at the same difficulty while the input grew longer. Across all its experiments, performance “consistently degrades with increasing input length.” Anthropic calls this an “attention budget”: each new piece of text uses some of it up. Its advice is to find “the smallest possible set of high-signal tokens” that gets the job done.
Order matters too. In a 2023 study led by Nelson Liu, models answering questions from a pile of documents did best when the useful document came at the beginning or the end, and much worse when it sat in the middle. I-PASS puts illness severity first for a reason.
What goes in can also steer the answer. Google says it designed its AI Overviews to show only what top web results backed up. Asked how many rocks a person should eat, the material it found included a satirical article, and the answer followed it. Google’s fixes changed what went in, limiting satire and user-generated content.
Handing over to yourself
Long jobs raise the same problem inside a single agent. When a conversation nears the limit, Claude Code summarizes it and starts a fresh window with the summary, keeping decisions and unresolved bugs and dropping old tool output. Anthropic warns that cutting too hard can lose “subtle but critical context whose importance only becomes apparent later.” Other agents keep running notes outside the window and read them back, or send helpers off to explore and take back only a short summary. Each is a handoff, sometimes from an agent to a later version of itself.
Two weeks away
Imagine a team lead about to take two weeks off, handing a client account to a colleague. The temptation is to forward every email thread and the project folder and call it done. Hundreds of messages arrive, and the one that matters, that the client’s legal team must sign off on any change to the contract, is buried in the middle of a thread from March.
Something closer to I-PASS works better. Start with what is urgent or fragile. Give a short summary of where things stand, then the list of what has to happen, then what could go wrong and what to do if it does. Link the full files for anyone who needs them. And before leaving, ask the colleague to say back, in their own words, what they will do first.
The same habits work with AI. Put the most important instruction first, and say why it matters. Hand over what the task needs, not everything you have. Point to where the rest lives, so the agent can fetch it when needed. Before a long job, ask it to restate its plan, and read that restatement as you would a colleague’s.
At those nine hospitals, the handoff wasn’t over when the doctor going home stopped talking. It was over when the doctor staying had said it all back.