At 3:27 on the afternoon of January 15, 2009, US Airways Flight 1549 was climbing out of LaGuardia Airport in New York. At about 2,800 feet above the ground, the captain, Chesley Sullenberger, said one word: “birds.” The cockpit recorder caught thumps, then a shudder. Both engines began to wind down.
Sullenberger and his first officer, Jeffrey Skiles, had not flown together until that week. According to the National Transportation Safety Board, this was the first trip the two had been scheduled to fly as a pair, four days that began on January 12. Skiles had more than 15,000 hours in the air, but only 37 in the A320, the airplane they were sitting in.
About 13 seconds after “birds,” Sullenberger said, “My aircraft.” He later told Congress how he split the work. He would fly, “since I had the greater experience in this particular aircraft type.” Skiles would run the checklist, “since he had so recently been through training on the airplane and knew the checklist intimately.” When the captain called for that checklist, about 17 seconds after the strike by his account, the first officer already had it out.
They had flown together for three days. They had about three and a half minutes.
A pickup game
Aviation had reason to think strangers were a risk. In a 1994 review of major US airline accidents in which the crew’s performance played a part, the NTSB found that 11 of the 15 with records had happened on the captain and first officer’s first day flying together, more than it expected from how airlines paired crews. It also warned that such accidents are rare and have many causes at once.
Flight 1549’s crew had a shared way of working instead. The NTSB found that “each pilot adhered to his role and responsibilities,” and that the captain “used the first officer as a resource by requesting his input.” Near the end he asked Skiles, “got any ideas?” The answer: “actually not.” Both pilots indicated to investigators that crew resource management, the cockpit teamwork training US Airways built into all its courses, was integral to the outcome. The captain, as the report summarizes him, said it gave pilots the skills “to build a team quickly.”
People who just met can still work as a team, if they share a playbook for roles and talk.
Amy Edmondson of Harvard Business School calls this teaming. “Teaming is teamwork on the fly,” she wrote in Harvard Business Review in 2012, “a pickup basketball game rather than plays run by a team that has trained as a unit for years.” Her book Teaming came out the same year.
What a stable team builds over months, teaming has to borrow. The rules of the game (who flies, who reads, how a handoff is said) come with the people. So does the habit of saying what you know and asking for what you don’t. Edmondson is blunt about what makes teaming hard: experts “often clash,” and when junior and senior staff from different divisions are paired, “reporting structures and hierarchies often silence dissent.”
It isn’t a case against stable teams. Edmondson writes that they “still work beautifully in many contexts,” and the NTSB numbers suggest familiarity still counts for something. Teaming is what you do when you can’t have it.
The call where nobody owns the customer
Imagine you run customer support at a software company, and a customer’s data export has gone to the wrong client. Within the hour you’re on a call with a security engineer, a company lawyer and the customer’s account manager. You have never worked with any of them.
The call can go one way. The engineer explains the bug, the lawyer lists obligations, the call ends on time, and each person leaves sure that someone else will phone the customer. Two days later, the customer hears about it from the client who got the file.
Or someone spends the first two minutes doing what the cockpit did in thirteen seconds. Who calls the customer, who confirms what leaked, who drafts the notice, said out loud and repeated back. The engineer says what she knows and what she doesn’t yet. You ask the account manager, who hasn’t spoken, what the customer was promised last month. The customer hears from one person, once, the same afternoon.
Agents that just met
AI agents get assembled the same way, a supervisor and a few specialists for one job. In 2025, researchers at UC Berkeley and elsewhere analyzed 1,642 recorded runs from seven multi-agent frameworks and catalogued where they broke down.
One run shows the problem in miniature. A supervisor agent held the user’s passwords. A phone agent held the phone app’s documentation, which said the username was the account’s phone number. The phone agent tried an email address, got “Invalid credentials,” and asked the supervisor for “the correct username and password for the phone app.” It never mentioned the phone number. The supervisor never asked. The service agents in that setup had been instructed to seek clarification. The authors read failures like this as agents failing “to accurately model other agents’ informational needs.”
You can look for the same gap. Read the messages your agents send each other, along with the final answer, and look for a specialist that reports a failure without the fact behind it. Then try a fixed handoff (what I tried, what I learned, what I need) and run the task several times with and without it. Check the transcripts against what the task actually required; one clean run is a lead, not proof.
In the cockpit of Flight 1549, by Sullenberger’s account, the handoff took four words. “My aircraft.”“Your aircraft.”