AI agents wrote 12× more PRs. Here is how to cut the CI bill in half
How much do AI coding agents add to a GitHub Actions bill?
When AI coding agents write your pull requests, CI becomes the bill that grows fastest. At jacks.med, the repository for our scheduling product went from 224 pull requests in May 2026 to 2,682 in the first 26 days of October, 12 times as many, and nearly all of them were written by Claude Code and Codex agents. Every one runs the full CI suite. At a median of 147 job-minutes per merged PR, October’s first 26 days come to about 370,650 runner minutes. At the 2 vCPU rate that’s about $2,224 at GitHub’s list price, $1,112 at Sengi’s list price, and $667 at the Sengi partner rate jacks.med pays. That’s 70% less than GitHub.
We run it ourselves: jacks.med is built by the same founders who build Sengi, so every number here is from our own bill.
Last checked October 27, 2026. Pull request counts from GitHub search on the jacks-med/onCall repository, measured October 26, 2026; CI minutes from a sample of 25 merged PRs, October 20 to 25, 2026. Every number on this page is in one JSON file.
How many pull requests do AI agents open?
Pull requests opened in jacks-med/onCall, from GitHub search:
| Month (2026) | PRs opened | PRs merged |
|---|---|---|
| May | 224 | 219 |
| June | 318 | 310 |
| July | 604 | 580 |
| August | 1,199 | 1,160 |
| September | 2,215 | 2,120 |
| October 1–26 | 2,682 | 2,518 |
183 of May’s 224 PRs carry Claude Code’s “Generated with Claude Code” line. In October, 1,740 carry it and 549 mention Codex. 1,113 of October’s 2,682 PRs were opened by the bot on our own agent farm, and 80 of the latest 100. The rest were opened from the founders’ accounts, mostly by agent sessions running on their laptops.
Where do the CI minutes go?
In a sample of 25 merged PRs from October 20 to 25:
- a median of 147 job-minutes per PR, and a mean of 153
- between 39 and 100 workflow runs on each of the last seven PRs we checked, because agents push several times while CI checks their work
What does that cost?
October so far: 2,518 merged PRs × 147.2 median minutes = 370,650 minutes.
| Runner | Label | Price per minute (2 vCPU) | 370,650 minutes, Oct 1–26 |
|---|---|---|---|
| GitHub | ubuntu-latest | $0.006 | $2,224 |
| Sengi, list price | sengi-standard-2-ubuntu-2404 | $0.003 | $1,112 |
| Sengi partner rate (what jacks.med pays) | sengi-standard-2-ubuntu-2404 | $0.0018 | $667 |
Sengi’s list price is half GitHub’s per minute, and the partner rate is 70% less, before counting faster builds.
Running agents at volume? If your team ships hundreds of agent-written PRs a month, ask for the partner rate: write to ops@sengi.run with your monthly minutes.
What did we change as agent volume grew?
All of these are true in our repository today:
- Cancel superseded runs. 12 of our 44 workflows use a
concurrencygroup withcancel-in-progress: true, so an agent’s fifth push stops the first four. - Path filters. 12 workflows run only when the files they check change.
- Route jobs by size and idle capacity. A small routing job sends each job to an idle 2 or 4 vCPU runner, either a desktop machine in our office or Sengi (29 workflows use it). Tiny jobs run on our own small machines.
- Database tests in CI, not on laptops. Agents push and let CI run the heavy suites instead of filling developer machines.
More changes, each with its YAML and its saving: how to reduce GitHub Actions costs.
Why is robust CI worth every penny when agents write the code?
Agents write code faster than anyone can read it. CI is what lets you trust it. In Vibe Coding, Gene Kim and Steve Yegge describe agents that quietly deleted or weakened tests to get them passing, and they argue for fast, frequent feedback loops as the thing that keeps AI-speed development stable (the book, The Register’s review). We track our agents’ work in Yegge’s Beads, the issue tracker built for coding agents and the work-tracking layer under his multi-agent orchestrator Gas Town.
What that looks like in our repository today:
- CI is the reviewer of record. Every merge needs green CI. For changes that touch production data, it also needs an independent adversarial review recorded as a required check on the exact commit.
- Red first. A bug fix lands its failing test as its own commit before the fix, so CI proves the test can fail.
- Robots for the boring parts. A lander workflow renumbers and lands database migrations, and a responder workflow picks up a red main.
- 44 workflows cover typechecks, database tests, end-to-end tests, security scans and deploys.
A broken production deploy costs more than a month of runner minutes. When CI is done right, it’s the cheapest insurance an agent-heavy team can buy, and on Sengi that insurance costs half as much.
Why do teams switch to Sengi?
- One line to switch. Change
runs-onand merge. Your workflows, secrets and logs stay in GitHub Actions. - Half the price, faster builds. $0.003 a minute at 2 vCPU against GitHub’s $0.006, on AMD Ryzen 9 7950X3D desktop CPUs boosting past 5 GHz, with jobs ready in 2 seconds.
- Partner rate for heavy users. $0.0018 a minute, plus priority routing, for teams running agents at volume.
- Real people behind it. Write to us and the founders answer. We work hard for every customer, and we’ll help you move your workflows over.
See how Sengi compares with every major runner provider in the runner comparison, and what agentic coding is worth to GitHub across all repositories in how much agentic coding makes GitHub in Actions dollars.
What is jacks.med?
jacks.med builds call, rotation and provider schedules for residency programs and attending groups. It generates each schedule with server-side optimisation solvers, so every rule is checked and everyone gets a fair share. The agent-written PRs above are how two founders ship it.
Where is the data?
Every number on this page, with its unit, period and source, is in /data/ai-coding-agents-github-actions-bill.json: pull requests per month, the CI-minute sample, October’s minutes and the cost at each per-minute price.
Frequently asked questions
Do AI coding agents increase GitHub Actions costs?
Yes. Every agent PR and every push runs your CI, and our PR count grew 12× in five months. Sengi is how you keep that growth from doubling the bill.
How many pull requests do AI coding agents open?
At jacks.med, 2,682 in the first 26 days of October 2026, against 224 in May 2026, nearly all written by Claude Code and Codex agents. 1,740 of October's PRs carry Claude Code's "Generated with Claude Code" line and 549 mention Codex.
How do I keep the bill down?
Cancel superseded runs, filter by path, size the runner to the job, and pay less per minute: Sengi is $0.003 a minute at 2 vCPU against GitHub's $0.006, and high-volume teams can ask for the $0.0018 partner rate.
What is the Sengi partner rate?
$0.0018 a minute at 2 vCPU, 70% less than GitHub's $0.006, plus priority routing, for teams running agents at volume. jacks.med pays it. Write to ops@sengi.run with your monthly minutes to ask for it.
Is CI worth the cost when agents write the code?
Yes. It is the check that catches an agent deleting or weakening tests, and every merge depends on it.
Do agents need a special runner?
No. They use the same workflows; only the runs-on line changes.
The fastest hardware for your GitHub Actions.
AMD Ryzen 9 7950X3D desktop CPUs boosting past 5 GHz, on bare metal with local NVMe, a fresh VM for every job, and 2 vCPU jobs ready in 2 seconds. Builds finish up to 1.9× faster than on GitHub’s runners, and at up to half GitHub’s price per minute they cost up to 74% less per build. Change one runs-on line, or give your coding agent one prompt and it opens the pull request for you.