Generated with costgrep v0.2.1 on 2026-10-02. Every case lists: the story, the numbers, links to the actual PRs and workflow runs, and the one command to reproduce. Sample = last ≤200 runs per repo (busy repos ≈ one day of traffic; window stated per case). List-price model — what this compute costs at GitHub's 2026 price list (see caveats in the benchmark). Evidence CSV per repo sits next to this file.
Repo: github/gh-aw — GitHub's product repo for Agentic Workflows. Window: 200 runs on 2026-10-02 (one day).
| who triggers CI | runs | share of runs | share of spend |
|---|---|---|---|
| AI agents (Copilot) | 107 | 53.5% | 28.8% ($1.15 of $3.97) |
| humans | 83 | 41.5% | 65.2% |
| bots | 10 | 5.0% | 6.0% |
Why run-share exceeds cost-share: most of the agent runs were still awaiting human
approval (action_required — 0 jobs executed, 0 compute). The agent creates the runs;
the minutes burn once they're approved.
See it yourself:
- Agent run (raw check): actions/runs/36993383799
— workflow "Addressing comment on PR #64994",
triggering_actor: Copilot, job 22m43s → 23 billable min × $0.006 = $0.1380 (matches our CSV row exactly). - The PR the agent was fixing: gh-aw#64994.
- A whole agent-driven workflow: "Running Copilot Code Review" in the repo's Actions tab.
Reproduce:
node scripts/costgrep.mjs --repo github/gh-aw --days 1 --max-runs 200 --csv-file gh-aw.csvRepo: Otisigma/halo-world-global-platform — a solo-maintained hobby project. Window: 200 runs, 2026-09-30 → 10-02.
| who triggers CI | runs | share of runs | share of spend |
|---|---|---|---|
| AI agents (Copilot) | 92 | 46.0% | 56.9% ($1.28 of $2.26) |
| humans | ~98 | 49.0% | 39% |
| bots | 10 | 5.0% | 4% |
Nearly half the pipeline runs and more than half the compute value in this repo is triggered by a coding agent, not by the maintainer. The workflow is literally named "Running Copilot cloud agent".
See it yourself:
- Agent runs: actions/runs/37001448241,
runs/36999980999 —
check the
triggering_actoron the run page (Copilot). - Agent-authored PRs: #278, #277, #276.
Bonus — agents work in serious repos too: zooba/pymsbuild (maintainer = a CPython core developer): 7 of 25 runs (28%) in the window triggered by Copilot, incl. PR #125 "Expose target Python and wheel-tag properties to MSBuild".
Reproduce:
node scripts/costgrep.mjs --repo Otisigma/halo-world-global-platform --days 3 --max-runs 200Repo: microsoft/vscode. Window: 200 runs on 2026-10-02 (one day). List-price value of the sample: $139.08.
| runner class | jobs | minutes | list-price $ | billed on a public repo? |
|---|---|---|---|---|
macOS 5-core M2 Pro (macos_xl) |
104 | 1,173 | $119.65 | YES |
| macOS standard | 23 | 64 | $3.97 | YES |
| Linux standard | 306 | 2,460 | $14.76 | no (free for public repos) |
| Windows standard | 22 | 71 | $0.71 | no |
| self-hosted | 248 | 2,194 | $0.00 | n/a |
89% of the compute value in this sample ($123.61) sits in runner classes GitHub bills even for public repositories. Public repos get Linux/Windows standard minutes free — macOS (and larger) runners are the exception, and in macOS-heavy projects they are the bill. This repo also shows 15 agent-triggered runs (7.5%) — agents and billed runners compound.
See it yourself: e.g. actions/runs/37000636285 ("Code OSS", 12 min on a 5-core macOS runner — $1.22 at list price), or open any recent run of the Code OSS build workflow and check the runner label.
Reproduce:
node scripts/costgrep.mjs --repo microsoft/vscode --days 1 --max-runs 200 --csv-file vscode.csv- Every repo has a per-job CSV next to this file — recompute
minutes × ratein any spreadsheet; the sums equal the headline totals (verified to the cent). - The attribution column is the API's own
triggering_actor.login, unmodified — click any run link above and compare. - Deterministic: same repo + window → same numbers. Rate table: GitHub's published 2026 list prices (source linked in every report's provenance block).
- Costgrep itself states the money caveats in every output: list-price vs invoice, public-repo freeness, rate-fallback jobs.