Ship PRs that speed up and cut cost of CI workflows
An agent loop on CI runners opens PRs that improve test and workflow speed and cost; it cut Mastra's slowest workflow from 29m56s to 5m06s and found a bottleneck in better-auth. ·
Done with a custom-built or unnamed agent
- What they did
- You install the StarSling GitHub App and switch your workflows to its runners by changing one runs-on line. Agents then read your workflows, jobs, run logs and machine telemetry, and open PRs that fix caching, shard tests, parallelize jobs and reorder work. A person reviews and merges each PR.
- How it went
- Mastra's test suite p95 fell from 29m56s to 5m06s. Better Auth's E2E went from 2m22s to 1m04s. Partcl says an agent PR made its Rust tests 2x faster. These figures are partly from the faster runners, not the agent PRs alone.
- Worth knowing
- Optimization PRs are only on paid plans and are off by default for new accounts. Usage is $0.008/min for a 4 vCPU runner, and the free trial gives 2,000 minutes.
Try it yourself with Codex
Look through the CI workflows in my repo [repo name or path] and find the slowest or most expensive ones using recent run times from [CI provider, e.g. GitHub Actions]. For the top [number] offenders, propose and make changes such as caching, parallelizing jobs, or skipping redundant steps. Open a separate draft PR for each fix that includes before and after timing estimates, and stop once each PR is ready for my review. Do not merge anything or change secrets or deployment settings without showing me first.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on HN · Nov 17, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Vibe-code and deploy an idle clicker game
Gilbert Pellegrom had his OpenClaw agent build a simple idle game from a conversation, commit it to GitHub and auto-deploy it to Vercel.
Test features on screen and merge fixed PRs
Ryan Carson uses Devin's Test App feature to record and narrate feature clicks, fix bugs, and merge PRs with a !land command.
Code with OpenCode and DeepSeek instead of Claude Code
A developer switched their coding agent from Claude Code to OpenCode with DeepSeek, reporting comparable quality at far lower cost for coding sessions.
Build an app from a simple prompt (partial result)
The poster asked Replit to build an app from a simple prompt, but the login did not work, and they hit the free usage limit while feeding in errors.
Build a trading bot with cursor-agent automation
A commenter is building a personal trading script/bot and uses cursor-agent with composer-1 to automate coding work.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.