The pitch at GitHub Universe was delegation: hand the agent a task, come back later, review the result. The pipeline that checks that result is a queue with a spinner on it. Both things were true in the same week, and engineers noticed.
We made the design for that gap.
What happened
GitHub Universe opened on October 28, 2026. The headline was Copilot moving from autocomplete to an agent that owns whole tasks, with a featured session titled "Stop prompting, start delegating".
The Octoverse report backed it with numbers. Developer Tech and TNW both report that more than 40% of pull requests in September started with an agent draft, out of about 110 million PRs. The report also says the bottleneck has moved from writing code to reviewing it. Tech Times asked in its headline whether a platform built in 2008 can handle that volume.
Meanwhile, October was busy on the status page. The incidents below come from GitHub Status and the unofficial incident summaries:
- Oct 7: two write-operation outages, about 19 minutes in total. They hit Git operations, pull requests, Actions, auth and Copilot. In the first, 17% of Actions runs failed with infrastructure errors.
- Oct 19: Actions, Pages and webhooks were degraded for about 67 minutes. An internal service lost its connections to cache servers and retried them all at once, which overloaded a shared network link.
- Oct 23: result processing slowed because a storage dependency throttled requests.
- Oct 26: a 12-minute API latency incident, also logged by UptimeRobot.
- Oct 30: another degradation of Actions, Pages and webhooks.
These were real incidents that affected real teams, so we're not here to point fingers. The joke is about the shape of the situation.
Why engineers get it
An agent changes the economics of opening a PR. Writing the diff used to be the slow part. Now it's nearly free, and the cost moves downstream: review, tests, builds, deploys.
Every one of those PRs wants a CI run. Multiply the PR count and you multiply the load on the one thing that decides whether the work is trustworthy. A queue doesn't care how fast the producer is.
The Oct 19 cause will sound familiar to anyone who has been on call. Something loses its connections, everything retries at once, and the retries become the outage. It's the same dynamic as a pile of agent PRs hitting a pipeline together: a polite, well-meaning crowd that jams the door.
Then there's the human at the end. Delegation sounds like less work, but the person who delegated still owns the review. They're now looking at a stack of green-pending checks, a yellow spinner and a status page that says "investigating". Nobody has built a good dashboard for "I trusted the robot and the robot's homework is waiting for a runner."
The lesson isn't that agents are bad or that CI is bad. Throughput is only as high as the narrowest stage, and the narrow stage moved.
The design
The design is a two-panel scene. At the top, a smug robot arm drops a huge stack of paper pull requests onto a desk. Below it is one tiny yellow queued spinner and a very tired engineer under a desk lamp. The printed line is "Agent opened 40 PRs. CI: still queued." The 40 is a nod to the 40% figure, not a measurement. The spinner is a generic one: no logos, no mascots, just a circle that won't finish.
Cheaper to Start. Costlier to Finish.
Git Push and Pray