In short6 min read
- StarSling took Mastra's slowest workflow from 29m 56s to 5m 06s, per the founders.
- Its agents test each change side by side and only ship statistically significant wins.
- New accounts get 2,000 free minutes; agent PRs start around 10,000 minutes.
- The propose-test-judge-ship loop is a reusable pattern for any automation you sell.
If you build with coding agents, your bottleneck has probably moved. Writing code got fast; waiting for tests did not. StarSling, a new Y Combinator company, is betting that the next problem for agent-heavy teams is continuous integration, and its launch post lays out both the numbers and the exact loop its agents run.
This breakdown covers what the founders built, why they built it, the results they shared, and a reusable pattern any solo builder can borrow, even if you never touch their product.
In this article
- The problem: agents ship PRs faster than CI can test them
- What StarSling actually sells
- The results the founders shared
- Inside the agent loop: five steps, one rule
- How the infrastructure decision played out
- How they are getting customers
- Should a small team care? A quick decision guide
- The pattern worth stealing: propose, test, judge, ship
- Before you try it: a short risk check
The problem: agents ship PRs faster than CI can test them
The founders, Yonas and Daniel, say they started building AI agents for DevOps last year because shipping code to production had fallen far behind agentic coding. Then they hit the wall themselves. According to the founders, their own agents shipped so many pull requests that at peak their job queues were backed up for hours, and their CI bill got close to their inference bill.
That last detail matters. Most people budget for model tokens and forget that every agent-generated PR also triggers a test run. When agents multiply the PR count, CI cost and queue time multiply with it.
The team has relevant background. One founder previously started StackShare, which grew to over 1.5M developers, and Daniel helped build Netflix’s internal developer portal.

What StarSling actually sells
StarSling is two things stacked together:
- Fast GitHub Actions runners. Each CI job gets its own ephemeral sandbox on Zen 5 machines with NVMe disks, with unlimited concurrency so a burst of agent PRs can all run at once. Billing is per second.
- Agents that keep improving your CI. A harness runs an agent loop tuned to make your workflows faster and cheaper, then sends the changes to you as PRs.
They also recently shipped StarSling Disks, persistent disks that the founders say give a 4.9x speedup in getting dependencies ready compared with actions/cache.
The results the founders shared
All figures below come from the launch post.
| Repo or test | Before | After / result |
|---|---|---|
| Mastra’s slowest workflow | 29m 56s on GitHub-hosted runners | 5m 06s on StarSling (5.87x) |
| better-auth test job | 11-month-old bottleneck | One job became 5.4x faster |
| benchmark-gogs | GitHub-hosted baseline | 12.8x faster |
| benchmark-transfer.sh | GitHub-hosted baseline | 11.7x faster |
| benchmark-mirotalksfu | GitHub-hosted baseline | 36.4x faster |
| benchmark-shellcheck | GitHub-hosted baseline | 5.5x faster |
| benchmark-chatwoot | GitHub-hosted baseline | 3.9x faster |
Across the open-source repos they benchmarked, the founders report an average 10x speedup in wall-clock time with 70% less runner time. For context, Mastra now merges over 1,800 PRs a month, which is exactly the kind of volume where minutes per run turn into real money.
The better-auth story is the more interesting one. The founders say they only realized the value of running an agent loop on top of fast runners when their agents found that 11-month-old performance bottleneck. Faster hardware alone would not have found it.
Inside the agent loop: five steps, one rule
The harness runs agents that:
- Analyze your workflows, run logs and machine telemetry
- Propose an optimization
- Test the unchanged workflow side by side with the proposed change
- Deterministically judge whether the difference is statistically significant
- Ship the winner as a PR with benchmark evidence
The rule hiding in step four is what makes this trustworthy: the agent does not decide on vibes. A deterministic check decides whether the change is real, and only then does a human see a PR. The founders sum up their lesson this way: fast infrastructure is table stakes for CI, but it is not sufficient on its own; you need a process that keeps making improvements.
How the infrastructure decision played out
The team started building runners on AWS at the beginning of the year and found it too expensive. At the time, they say, no other cloud provider offered the same machines billed per minute as microVMs. So they convinced a handful of sandbox providers to get Zen 5 machines and began benchmarking them publicly.
For anyone building an infrastructure-heavy product, that is a useful reminder: the default cloud is a starting point, not a requirement, and publishing your benchmarks doubles as marketing.
How they are getting customers
The go-to-market moves in the launch are worth studying even if you sell something completely different:
- An agent-first onboarding. Instead of a signup form, you give your coding agent a docs URL and it sets things up. That fits an audience that already lives inside agents.
- A free tier sized to the product’s threshold. New accounts get 2,000 free minutes, and the agents start opening PRs once you have hit about 10,000 minutes. The free tier lets you test the runners; the optimization loop needs real usage data.
- Giving the method away. The founders acknowledge you can do these optimizations yourself with prompts, so they released their skills publicly, plus CI playbooks based on practices from OpenAI, Anthropic, Linear and others. Their pitch is that their agents save you the tokens and CI compute of finding optimizations yourself.
- A risk-reversal offer. Their CI Challenge says that if StarSling cannot save you at least 10% on your next CI bill, they will give you $1,000, subject to the eligibility rules on their site.
Should a small team care? A quick decision guide
This is our own framework, not StarSling’s. Most solo founders do not have a CI problem yet. Use these signals to judge whether you do.
| Signal in your setup | What it suggests |
|---|---|
| Agents open several PRs a day and runs queue up | CI speed is now your shipping bottleneck; worth testing faster runners |
| CI spend is creeping toward your model spend | The founders’ own warning sign; audit workflows before scaling agents |
| A few runs a week, under your included minutes | Not worth switching; borrow the playbooks instead |
| Tests are slow but you cannot explain why | An analysis pass (manual or agent) likely beats new hardware |
| You build automations for clients | The loop pattern below is the real takeaway |
If you already use agents like the ones in our Baloon review on turning Jira tickets into PRs or the Momentic QA agent review, PR volume is exactly where this cost shows up first.
The pattern worth stealing: propose, test, judge, ship
The StarSling loop is not really about CI. It is a template for any agent you let near production, including the n8n, Make or Zapier automations many freelancers sell.
A typical failure with AI agents in client work is letting the agent both make a change and declare it a success. StarSling splits those roles. Applied to your own automation work, that looks like this:
- Collect evidence first. Logs, run history, error rates. The agent proposes nothing until it has read them.
- Run old and new side by side. Keep the current workflow live and run the proposed version on the same inputs.
- Let a deterministic check decide. A fixed threshold or test suite, not the model’s opinion, says whether the change is better.
- Ship as a reviewable proposal. A PR, a diff or a short report with before-and-after evidence, so a human approves it.
That structure also makes automation services easier to sell, because every change arrives with proof. Our hub on AI agents for solo businesses collects more examples of keeping agents on a short leash.
Before you try it: a short risk check
- Results are theirs, not yours. The speedups above are the founders’ benchmarks on specific repos. Your gains depend on how your workflows are built.
- The loop needs volume. Agent PRs start around 10,000 minutes, so a light user will mostly be testing runners, not the optimization agents.
- Read the challenge rules. The $1,000 offer has eligibility conditions; check them on the official site before counting on it.
- Review every PR. Even with statistical checks, an agent changing your test setup deserves a human read, especially around caching and skipped tests.
- Pricing details. Per-second billing is confirmed, but check the official site for current rates before comparing against GitHub-hosted runners.
Time your slowest workflow before changing anything
Open your GitHub Actions history today and write down the duration of your three slowest workflows and how often they run. That baseline tells you whether faster runners, an optimization pass or nothing at all is the right next move.
Source: Launch HN: StarSling (YC P25) – Self-improving CI for agent-volume repos
Get new AI automation playbooks by email
One email when new guides and case studies go up. No spam, unsubscribe anytime.

Leave a Reply