知Yizhi Daily
← Back to this edition

Conversations

StarSling Launches GitHub Actions Runners with AI Optimization Agents

Edition 2026-11-18 · 08:00 Beijing time

YC P25 team StarSling posted on Hacker News a CI service aimed at repositories with high-frequency agent PRs. Its GitHub Actions runners provide ephemeral Zen 5 sandboxes, NVMe disks, and unlimited concurrency, while AI agents continuously submit PRs for test and workflow optimizations.

Founders Yonas and Daniel say StarSling's CI harness analyzes workflows, run logs, and machine telemetry, proposes optimization plans, then tests them side by side against unmodified workflows to determine whether differences are statistically significant, and finally submits the winning plans as PRs along with benchmark evidence. Customer Mastra merges more than 1,800 PRs per month, and its slowest workflow dropped from 29 minutes 56 seconds on GitHub-hosted runners to 5 minutes 06 seconds, a 5.87x improvement; the agent also found an 11-month-old performance bottleneck in better-auth, speeding up one test task by 5.4x.

The service supports per-second billing and has introduced StarSling Disks persistent disks, which the company says prepare dependencies 4.9x faster than actions/cache. For benchmarks on open-source repositories, the team claims an average 10x wall-clock speedup and a 70% reduction in runner time, with example projects improving by factors ranging from 3.9x to 36.4x. The team has also published skills and CI playbooks that can be used in prompts, but its commercial selling point is that agents bear the token and CI compute costs required to find optimizations.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader