The Human Context Window Is Your New Bottleneck

A good developer on one of my client teams now runs three to six tasks in a single day. One agent refactors the billing module while another writes tests for last week's feature. A third chases a flaky integration. The developer, meanwhile, is reviewing a pull request with all of that churning in other terminals.

That's normal now, and the output is real. I've watched teams ship in a month what used to take several. There is a cost nobody puts on the dashboard, though, and I've started hearing it in people's voices on the Monday call.

The agent has a huge context window. You have a TI-82.

Models hold hundreds of thousands of tokens of context. A whole codebase fits in view. The person directing them is working with something closer to a TI-82 graphing calculator: a handful of things held at once, and everything else written to disk, badly.

Psychologists put working memory at roughly four meaningful chunks. Every running agent costs at least one of those. You're tracking what it's doing, what you told it, where it's likely to go wrong and what you'll need to check when it comes back. Run five tasks and you've spent the budget before opening Slack.

The part that surprised me is which work AI took away. Writing a function you've written a hundred times used to be low-load work; your hands did it while your head idled. Agents ate that. What's left is the expensive kind: deciding, reviewing, catching the subtle thing the model got confidently wrong. A developer running six agents lives in that mode from nine to six.

Switching got cheap to start and expensive to carry

Context switching used to have a natural brake. Starting a second task meant putting down the first, so people rarely juggled more than two. Now a new task costs one prompt. The brake is gone and the load stacks.

Each switch leaves what researchers call attention residue: part of your head stays on the thing you just left. With agents, the thing you left keeps moving. It finishes, or it gets stuck, or it wanders somewhere you didn't want, and some background process in your brain is waiting to find out which. Multiply that by five and you have a person who is never fully present on anything.

It shows up in mood first

Output is a lagging indicator. By the time velocity drops, someone has been running hot for weeks. The early signals are softer, and a manager who only watches the board will miss every one of them.

• Shorter fuse in code review. Comments get curt, and small disagreements turn into long threads.
• Late-night and weekend commits from someone who never used to work late.
• A flat "I'm fine" at standup.
• More reverts, and bugs a rested version of the same person would have caught.
• Quietly dropping the optional stuff: demos, pairing, banter in the team channel.

None of that shows up in your DORA metrics for a while. It shows up in how someone sounds on a call. If your developers run multiple agents a day and nobody has asked them how they're actually doing this month, ask this week, one on one, and mean it.

Where this series goes

Part 2 is the policy I run with clients: extra days off after heavy weeks, averaging three or four a month, and why the math favors the employer. Part 3 is the operating manual, because a badly run version of that policy becomes a brand new way to burn people out.

If your team's output doubled this year and nobody has checked on the people producing it, let's talk →