StarSling, a Y Combinator-backed company, launched a continuous integration service that pairs fast GitHub Actions runners with AI agents. The agents study a customer's workflows and submit pull requests to make tests and builds faster and cheaper over time.

Each CI job runs in its own ephemeral sandbox on Zen 5 machines with NVMe disks, according to the company. StarSling says it allows unlimited concurrency, so a burst of agent-generated pull requests can all run at once. That matters for teams whose code is increasingly written by machines. Co-founders Yonas and Daniel wrote in their launch post that their own agents shipped so many pull requests that job queues backed up for hours at peak, and the CI bill came close to the company's inference bill.

The product works through a loop. The agents analyze workflows, run logs and machine telemetry, then propose an optimization. They test the unchanged workflow side by side with the proposed change and judge whether the difference is statistically significant. The winning change ships to the customer as a pull request with benchmark evidence attached.

Mastra, which the company names as a customer, merges more than 1,800 pull requests a month. StarSling says it cut Mastra's slowest workflow from 29 minutes and 56 seconds on GitHub-hosted runners to 5 minutes and 6 seconds, a 5.87x speedup. The company also points to better-auth, where its agents found a performance bottleneck that had gone unnoticed for 11 months. The fix made one of that project's test jobs 5.4x faster.

StarSling also ran its system on several open source repositories. It reports an average 10x speedup in wall-clock time and 70% less runner time across those projects. The largest gain was on the mirotalksfu repository, at 36.4x. The gogs repository improved 12.8x, transfer.sh 11.7x, shellcheck 5.5x and chatwoot 3.9x. The benchmarks are published on GitHub under the company's demo organization.

The founders have ties to developer tools. Yonas previously founded StackShare, which grew to more than 1.5 million developers. Daniel helped build Netflix's internal developer portal. The company said it began building AI agents for DevOps last year, after it saw that shipping code to production had fallen behind agentic coding.

Infrastructure was a hurdle from the start. StarSling first built its runners on AWS at the beginning of the year but found the costs too high. At the time, no cloud provider offered the same machines billed per minute as microVMs. The company persuaded several sandbox providers to offer Zen 5 machines and has published its benchmarks of those providers on GitHub.

StarSling has added per-second billing and a new product called StarSling Disks, which are persistent disks. The company says the disks made dependencies ready 4.9x faster than actions/cache. It has also turned CI optimizations from OpenAI, Anthropic, Linear and others into playbooks that agents can follow.

New users get 2,000 free minutes. According to the company, agents begin opening pull requests once an account reaches about 10,000 minutes. Under the CI Challenge, StarSling will pay $1,000 to any customer whose next CI bill is not reduced by at least 10% through the service. Eligibility rules are posted on the company's blog.

Developers who want to make the changes themselves can use skills the company has released on GitHub. StarSling says its agents spare users from spending tokens and CI compute on the search for optimizations. The founders are asking users for feedback on the headless experience and on what they would like improved in their own CI setups.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.