StarSling is an AI-native CI platform for R&D teams whose GitHub Actions pipelines face heavy build load, long waits, and complex test workflows. Its goal is not only faster CI, but CI that keeps improving itself: analyze runs, propose optimizations, test them safely, and turn successful changes into production-ready pull requests. To make that loop practical, StarSling uses Novita Agent Sandbox as the execution layer for isolated microVMs, elastic compute, agent experimentation, and reproducible evaluation.
That combination matters because CI optimization is experimental by nature. A change to caching, parallelization, test sharding, or workflow structure can help one repository and hurt another. StarSling needs a way to try changes against real conditions without risking customer systems. Novita Agent Sandbox gives each experiment a controlled workspace, then leaves the promotion decision to evidence and review.
In this case study, you’ll see why CI optimization is difficult to automate, how StarSling runs its agent loop, what isolated execution changes, and how the same architecture applies to other AI workflows.
Why CI optimization is hard to automate
Many engineering teams already know what slow CI costs: delayed merges, idle review cycles, slower feedback, and less time for product work. The usual fixes are also familiar. Cache dependencies, shard tests, parallelize jobs, reduce redundant work, remove flaky checks, and restructure pipelines so important feedback arrives earlier.
The difficult part is not naming those ideas. It is proving which change is safe and effective for a particular repository. A workflow file encodes historical constraints, organizational preferences, timing assumptions, and dependencies that may not be obvious from logs alone. Two teams with similar build times may need different fixes. One may need better caching; another may need test isolation, more parallelism, or a smaller critical path.
Manual optimization also does not scale well across many repositories and customers. Someone has to inspect logs, form a hypothesis, edit the workflow, rerun jobs, compare results, and repeat. That process consumes expert time and makes it harder to preserve what worked as pipelines evolve.
This is where AI agents can help, but only if they have a safe execution boundary. StarSling’s agents inspect workflows and telemetry, generate optimization changes, and test those changes before proposing them. Running that loop directly inside production CI would make experimentation risky and hard to control. An agent needs to act, fail, retry, and measure without becoming an operational hazard.
How StarSling runs the optimization loop
StarSling treats CI as a system that can observe and improve itself. The workflow starts with evidence: build times, test behavior, logs, resource usage, and pipeline structure. From that evidence, StarSling can identify candidate bottlenecks such as repeated setup work, poorly sharded tests, unnecessary serialization, or workflow steps that spend more time than they should.
The next stage is experimentation. StarSling produces a proposed optimization and checks whether it actually improves the pipeline. Does it reduce wall-clock time? Does it lower cost? Does it preserve correctness? Does it fail cleanly when conditions change? Those questions require execution, not just static analysis.
Once an experiment succeeds, the result can become a concrete change for review: a workflow update, a cache adjustment, a test-sharding change, or another pipeline improvement. Keeping that output in a pull request gives engineers a clear artifact to inspect before it reaches production.
StarSling reports that customers have seen CI become up to 6X faster and up to 13X cheaper. Those are customer results, not universal guarantees; outcomes depend on repository shape, test suite, infrastructure, and the types of bottlenecks present. Still, the pattern is useful: pair automation with measurement so improvements are supported by observed pipeline behavior rather than guesswork.
What Novita Agent Sandbox provides
Novita Agent Sandbox gives StarSling isolated microVMs for agent workloads. Instead of sharing execution directly with customer CI systems or internal services, each agent run can have its own controlled environment. That separation reduces the chance that exploratory commands, dependencies, or failed experiments interfere with other work.
The second property is elasticity. CI optimization experiments are bursty: there may be many runs during investigation and fewer once a change is ready for review. Sandboxed compute lets StarSling run experiments when needed without making every customer implementation wait on constrained internal resources.
The third property is reproducibility. Reproducing an experiment helps distinguish a real improvement from noise. With an isolated environment, StarSling can evaluate agent actions under consistent conditions, validate changes, and decide when an optimization is ready to move forward with more confidence.
Together, these properties turn the agent loop into a practical product workflow rather than an ad hoc script. The sandbox is not where the human decision disappears. It is where candidates can be tested safely enough to make a useful decision.
Results for AI-native CI
The approach produces three connected benefits.
Faster optimization experiments. StarSling can shorten the time between identifying a bottleneck and shipping a fix. Instead of waiting for manual investigation cycles, candidate changes move through a structured loop of proposal, execution, measurement, and review.
Stable agent execution. Isolated environments reduce operational risk as agent-based automation scales. Agents can run repeatedly and fail safely, which is essential when the workflow itself is responsible for testing changes to CI.
Easier multi-customer scale. Novita Agent Sandbox lowers the barrier to deploying AI optimization into real CI pipelines. StarSling can support more customers without rebuilding the execution foundation for every new repository or use case.
Daniel Worku, StarSling CTO, described the operational effect:
“Novita Sandbox gave us the isolated environment to run CI workloads safely and reproducibly at scale. Their engineers are world-class and ship everything that we ask for in days not weeks.”
That quote points to the broader lesson: agent infrastructure has to be reliable before agentic automation can be useful. Isolation, elasticity, and repeatability are not implementation details; they are the conditions that let AI workflows operate close to production systems.
How to apply this architecture
You can reuse this pattern for any agent workflow that needs to change real systems safely.
First, separate the analysis plane from the execution plane. Logs and metrics can inform a candidate change, but the change should be tested in a bounded environment before it reaches production.
Second, define the artifact under review. In CI optimization, that is usually a workflow diff or configuration change. In code generation, it may be a patch. In data work, it may be a query, notebook, or report. A reviewable artifact makes the agent’s work auditable.
Third, measure before and after. For CI, track duration, cost, failure rate, flaky tests, queue time, and time to feedback. A single speed number can hide regressions elsewhere.
Fourth, keep permissions narrow. An optimization agent does not need unrestricted network access, production secrets, or permission to merge. Give it the minimum access required to run the experiment and produce evidence.
Fifth, keep a human gate for promotion. A sandbox can verify that a change behaved as expected under test conditions. Humans and existing review processes should still decide whether that change is appropriate for production.
FAQ
What is StarSling?
StarSling is an AI-native CI platform. It analyzes CI workloads, proposes improvements such as caching, parallelization, and test sharding, and generates optimization pull requests for engineering teams.
How does StarSling use Novita Agent Sandbox?
StarSling uses Novita Agent Sandbox to run agent workflows in isolated microVMs, access elastic compute for repeated experiments, and evaluate CI optimization changes in reproducible conditions.
Does the 6X speedup apply to every CI pipeline?
No. StarSling reports that some customers have seen CI become up to 6X faster and up to 13X cheaper. Results depend on the pipeline, test suite, repository structure, and bottlenecks involved.
Why do CI optimization agents need a sandbox?
A sandbox gives agents a place to run commands and test changes without directly interfering with customer CI systems or shared infrastructure. It also supports cleaner comparisons between experiments.
Recommended articles
- Run Codex or a Coding Agent in a Secure Sandbox
- Open Source AI Agent Sandbox Options: What to Check Before Self-Hosting
- Building a Coding Agent with Novita’s Agent Sandbox
Sources
- StarSling, “CI Harness v2: The Agent Loop That Shipped a 5.4x Speedup” — checked September 17, 2026: https://starsling.dev/blog/ci-harness-v2-the-agent-loop-that-shipped-a-5-4x-speedup
- Novita Agent Sandbox overview — checked September 17, 2026: https://novita.ai/docs/guides/sandbox-overview
