The day my agents repainted my GitHub year.

Tuesday evening, I drop by my GitHub profile. On the far right of the graph, there’s a dark green square. I hover over it: “618 contributions on October 10th”. That was Saturday.

The rest of the year has faded. Months of work, days I was quite happy with, all turned the same washed-out green.

Since late September, I’ve been running a fleet of AI agents across a dozen repositories, and they commit a lot. Their best day became the reference point of my graph, and everything I did before got squashed underneath it.

How one day repaints a year

I assumed GitHub colored squares against a fixed scale. In fact, it takes your highest day, cuts the range into four equal steps and files each day into one of them (developers checked this by recomputing the graph, on Stack Overflow). When I pulled the data a few hours later, my maximum was 622. So the palest step ran from 1 to 155 contributions, while my median active day has 17.

Before the agents With the agents

1 day stands out from pale green, of my 248 active days before the agents

GitHub splits the scale, from 0 to the highest day, into four equal steps. With a maximum of 622, any day under 156 contributions is pale green.

On September 26, my record was 184, and 39 of my 248 active days stood out from pale green. On October 13, only one is left. Those days haven’t moved by a single commit; the scale just climbed to 622.

What those 622 contributions are worth

That Saturday, my agent journal saw 541 commits. It’s a small collector running on my Mac that logs my sessions with the agents, my messages and the commits. Of those 541 commits, 487 are co-authored by an agent. GitHub also counts pull requests, reviews and issues, hence the gap with 622.

So that dark green square is mostly my agents’ work. And I’m probably not the only one: GitHub received one billion commits in all of 2025, and in April 2026 its COO announced 275 million… per week (Kyle Daigle, quoted by Simon Willison).

And where am I in all this?

To find out, I compared my 17 days of work with agents from September 7 to 25, before the fleet (the 2 days without any session don’t count), with the 16 days from September 27 to October 12.

Since my agents started committing, average per active day (before = ×1)

×1×2×4×6×8

• Commits41.8 → 285.1 per day

  ×6.8

• My time with the agents5.1 → 9.4 hours per day

  ×1.8

• My project switches11.8 → 77.9 per day

  ×6.6

Commits went up 6.8×, which I expected. My time with the agents nearly doubled, from 5 to over 9 hours a day, and my project switches went up 6.6×, in step with the commits. And those hours only count the quarter-hours in which I wrote to an agent: time spent reading, testing or thinking without typing isn’t in there.

Before the fleet, I moved from one repository to another every 26 minutes on average. Now it’s every 7 minutes.

Pratima Arora (Smartsheet) puts it well on the Stack Overflow blog: “The hours haven’t changed, but the density of work has” (Stack Overflow, May 2026). For me it’s worse: the hours nearly doubled, and on top of that the density more than tripled.

Saturday, hour by hour

Between 18:00 and 19:00, I switched projects 23 times, on average every 2 minutes and 36 seconds. Over the whole day, from 12:29pm to 1:38am, that adds up to 151 switches across 13 repositories. After my last message, the agents kept going on their own and pushed another 46 commits between 2am and 5am.

It reminds me of a study by Gloria Mark (University of California, Irvine) on interrupted work. People compensate by working faster, “but this comes at a price: experiencing more stress, higher frustration, time pressure and effort” (CHI 2008). In 2008, the interruption came from an email or a colleague. That Saturday, I interrupted myself 151 times, every time an agent had finished and was waiting for an answer.

The evening fog

This fatigue now has a name, “AI brain fry”, coined by BCG. In their survey of 1,488 US workers, 14% of those using AI said they had experienced it, and the most exposed oversee several agents at once (Axios, March 2026, on the study published in Harvard Business Review). They describe a buzzing in the head and decisions that take longer.

On GitHub, that day is still my best square of the year.

What I’d like to see instead

I had already measured my focus blocks in The apple tree and the basket: 13 minutes on average, against 81 a year earlier. What I was missing was a counter for context switches.

That’s why I’m building Humanload. It counts, on your machine, what GitHub doesn’t see, like the decisions you make for your agents or the sessions waiting for your answer. The numbers in this post come from its version zero, which is simply my own journal. If this speaks to you, the waitlist is open.

As for my GitHub graph, it will stay pale green at least until October 2027, when that Saturday leaves the one-year window. Or longer, if my agents beat their record before then.

Method. Contributions come from my public GitHub profile, pulled on October 13, 2026 (Paris time). Everything else comes from my agent journal, with days counted from 5am to 5am. I count a project switch when two of my consecutive messages go to two different repositories, and time with the agents in quarter-hours in which I sent them at least one message. All repositories that Saturday are personal.