StarSling Cuts Mastra's Slowest Build from 29 Minutes to 5
When AI agents send pull requests in batches, CI can't keep up: queues stretch for hours, and the CI bill catches up with the model bill. StarSling (YC P25) runs every check in its own sandbox on Zen 5, and its agents propose fixes based on the logs. Mastra's slowest build dropped from 29 min 56 s to 5 min 06 s (5.87x). If you have a lot of PRs, measure your CI first.
AI-processed from Hacker News AI; edited by Hamidun News
Coding agents have learned to open pull requests (PRs, proposed changes to code) faster than CI can check them. CI is the set of automated builds and tests that GitHub Actions runs on every PR according to a script called a workflow, and servers called runners execute them. The StarSling founders ran into this themselves: their agents sent so many PRs that job queues stood for hours at peak, and the CI bill came close to the inference bill, meaning the cost of running the models.
The product grew out of that problem, and it collected 81 points and 39 comments on Hacker News. The headline number from the launch: the slowest workflow at Mastra dropped from 29 minutes 56 seconds to 5 minutes 06 seconds, a 5.87x speedup.
That result comes from StarSling runners with AMD Zen 5 processors, compared with standard GitHub runners.
What the product is
StarSling is part of YC P25, the summer 2025 batch of the Y Combinator accelerator. The founders are Jonas, who previously created StackShare (which drew more than 1.5 million developers), and Daniel, who helped build an internal developer portal at Netflix. The product has two parts.
The first part is fast runners for GitHub Actions. Each CI job gets its own disposable sandbox on Zen 5 machines with NVMe disks. There is no concurrency limit, so a batch of agent PRs starts at once instead of waiting in a queue. Billing is per second. The company recently added StarSling Disks, persistent disks on which, according to the authors, a project's dependencies are prepared 4.9x faster than with the standard actions/cache.
The second part is agents that gradually make CI itself cheaper. They work in a loop:
- they analyze your workflows, run logs, and machine telemetry;
- they propose an optimization;
- they run the old and modified workflows at the same time to compare them;
- they check deterministically that the difference is statistically significant rather than random;
- they send the winner as a PR with benchmarks.
Put simply, your team gets an engineer who regularly looks for where the build wastes time and brings measurements instead of guesses.
What numbers are claimed
At Mastra, which merges more than 1800 PRs a month, the slowest workflow got 5.87x faster. The second example is better-auth: the agents found a performance bottleneck that had lived there for 11 months. The fix sped up one of the test jobs by 5.4x (PR #10762 in the better-auth repository).
The startup also ran five open-source repositories on its runners: gogs (12.8x), shellcheck (5.5x), chatwoot (3.9x), transfer.sh (11.7x), and mirotalksfu (36.4x). On average, that came to a speedup of about 10x in wall-clock time and 70% less runner time. The spread from 3.9x to 36.4x shows that results depend heavily on the project, and "10x" should not be taken as a guarantee.
The infrastructure is interesting too. The team first built the runners on AWS, but it turned out too expensive. Providers did not yet offer the micro-VMs they needed, with per-second billing and Zen 5, so the team persuaded several sandbox vendors to add them and published benchmarks publicly in the hpc-sandbox-benchmarks repository. The authors' conclusion: fast hardware is necessary for CI but not sufficient, and you also need a process of continuous improvement. So they turned CI speedup techniques from OpenAI, Anthropic, Linear, and other companies into playbooks for agents.
What to do with this and where the catch is
All the figures come from the authors themselves, and the benchmarks were run on a StarSling demo organization. The article offers no independent verification, so measure your own workflows before switching.
There is also an entry condition. The free tier gives 2000 minutes, but the agents only start opening PRs after roughly 10,000 minutes have been used. A small repository with infrequent builds will have to wait a long time for the "self-improvement" effect.
If you don't want to pay for the agents, StarSling has published a set of skills and playbooks with CI optimizations. You can hand them to your own agent. The authors themselves admit you can do this, but then you pay for the tokens and compute spent searching for optimizations. As a test, the startup offers a bet: if it doesn't save you at least 10% on your next CI bill, it will pay $1000 (the terms are described in the blog).
The practical takeaway is simple. If your team works with agents and the PR flow is growing, CI stops being background noise and becomes a line item in the budget. Look at how much time and money goes into downloading dependencies and repeated steps, then choose: adopt the fast runners, apply the playbooks yourself, or first just take baseline measurements.
- Nov 18, 2026AI Agents Shipped 228 Game Releases Without Human Review; Tests Sent Back 25% of Tasks
- Nov 18, 2026JetBrains' Mellum Runs Locally in an IDE: 5 Out of 10 on a MacBook M3 Pro
- Nov 18, 2026OpenAI Releases Legal AI: 95% Fewer Fabricated Facts Than Claude Opus 5.5
- Nov 18, 2026Claude SEO Plugin Runs Up to 17 Agents and Audits a Site with One Command
Key Questions Answered
What the product is?
StarSling is part of YC P25, the summer 2025 batch of the Y Combinator accelerator. The founders are Jonas, who previously created StackShare (which drew more than 1.5 million developers), and Daniel, who helped build an internal developer portal at Netflix. The product has two parts. The first part is fast runners for GitHub Actions. Each CI job gets its own disposable sandbox on Zen 5 machines with NVMe disks. There is no concurrency limit, so a batch of agent PRs starts at once instead of waiting in a queue.
What numbers are claimed?
At Mastra, which merges more than 1800 PRs a month, the slowest workflow got 5.87x faster. The second example is better-auth: the agents found a performance bottleneck that had lived there for 11 months. The fix sped up one of the test jobs by 5.4x (PR #10762 in the better-auth repository). The startup also ran five open-source repositories on its runners: gogs (12.8x), shellcheck (5.5x), chatwoot (3.9x), transfer.sh (11.7x), and mirotalksfu (36.4x).
What to do with this and where the catch is?
All the figures come from the authors themselves, and the benchmarks were run on a StarSling demo organization. The article offers no independent verification, so measure your own workflows before switching. There is also an entry condition. The free tier gives 2000 minutes, but the agents only start opening PRs after roughly 10,000 minutes have been used. A small repository with infrequent builds will have to wait a long time for the "self-improvement" effect.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.
