Before dawn on 21 August 2017, the destroyer John S McCain was overtaking slower ships in the Singapore Strait. One sailor at the helm was handling both the steering and the speed of the ship’s two propellers. About 5:20 the commanding officer, worried the helmsman “might become overwhelmed” in the crowded lane, ordered a second sailor to take over the propellers.
The split ran through a touch screen, one propeller at a time. Moments later the helmsman reported that steering had failed. It hadn’t: control of the rudder had shifted to the second station, and none of the three sailors handling the controls recognized it. The National Transportation Safety Board later found “a broad lack of knowledge and understanding” of the system among the crew. About three minutes after the first report, a tanker struck the destroyer, and 10 sailors died. The case has the rest.
How much a team can hold
A team can keep only so much in mind. Pile on more, and mistakes and delays tend to rise.
The idea starts with individual minds. In 1988 the educational psychologist John Sweller argued that the usual way of solving a problem, steadily closing the gap between where you are and the goal, takes “a relatively large amount of cognitive processing capacity,” leaving little for learning how such problems work.
Matthew Skelton and Manuel Pais carried the idea over to teams in their 2019 book Team Topologies. Their site puts it plainly: “Each new tool, responsibility or domain your team is given taxes their mental bandwidth,” and competent teams become ineffective “when leaders keep adding to their plate without taking anything away.” The consultancy Thoughtworks put “team cognitive load” on its Technology Radar in 2021, as a way to check how easy or hard teams find building, testing and maintaining what they own, using an assessment template based on the book’s ideas.
Load depends on what you know
The McCain shows why the same job can be light for one team and crushing for another. Splitting steering from thrust was meant to lighten the helmsman’s load. But on this ship, the split was itself a task: a touch-screen transfer, done separately for each propeller, in a steering mode where any station could take the rudder without anyone at the helm doing anything. The NTSB called the throttle transfer “unnecessarily complex.” The second sailor had come over temporarily from a cruiser without the system and told investigators he hadn’t known steering could be shifted to his station.
Fatigue shrinks capacity too. The 14 people on the McCain’s bridge had averaged just over 4.9 hours of rest in the previous day. And the load had been piling up for a while: two weeks after the collision, the Government Accountability Office told Congress that 37 percent of warfare certifications for cruiser and destroyer crews based in Japan had expired, and that across the Navy some sailors were working over 100 hours a week.
So load isn’t just a count of things a team owns. It depends on how familiar those things are, how well they are designed, and how tired the people are.
Where the idea is thin
Sweller’s research is about individuals learning; applying it to whole teams is Skelton and Pais’s extension, not something the 1988 paper tested. The assessment template Thoughtworks used is a questionnaire teams answer about themselves, which shows how loaded a team feels, not how close it is to failing. And taking things off one team’s plate often puts them on another’s.
Imagine a four-person IT team at a small company that runs the payroll system, the website, the office network and 20 other tools. Every request lands somewhere in that pile. Each one is easy alone, and the team may still answer slowly, because every call starts with remembering how that particular system works.
Tools for agents
AI agents have a hard limit on how much they can take in at once, their context window. In a 2025 guide to writing tools for agents, Anthropic warned that “more tools don’t always lead to better outcomes” and that “too many tools or overlapping tools can also distract agents from pursuing efficient strategies.” It recommends tools with a “clear, distinct purpose,” sometimes combining several operations into one, so fewer tool descriptions and intermediate results take up the agent’s context. That, the company wrote, “reduces an agent’s overall risk of making mistakes.”
The NTSB’s recommendation for the Navy pointed the same way: build complex systems to a standard that calls for simplicity. On the McCain, a change meant to lighten one sailor’s load added a task few on the bridge fully understood.