Are AI Coding Agents the New CI Bottleneck?

AI coding agents are moving the software bottleneck from writing code to validating it. Here’s what that means for CI infrastructure, capacity planning, and software quality.

This topic is explored in more detail in “CI Can’t Keep Up With AI,” an episode of Uplink, where Aditya “JP” Jayaprakash, Co-founder and CEO of Blacksmith, discusses how AI coding agents are reshaping CI and software delivery.

AI coding agents can move from a prompt to a pull request in minutes. GitHub’s own workflow guide shows an agent turning an issue into a tested, review-ready pull request while the developer gets to work on something else. That sounds like a productivity win — until each new pull request joins a queue of builds, tests, security checks, and reviews.

The reality is that code generation is getting faster, while code validation still has to prove that every change is safe. That makes continuous integration infrastructure one of the next constraints in AI-assisted software development.

What is continuous integration, and why does it matter?

Continuous integration (CI) is the process of automatically building and testing new code before it is merged into a shared codebase. It catches regressions early and gives developers confidence that a change hasn’t broken something that already worked.

As a codebase grows, that safety net becomes heavier. Even a small change can trigger thousands of tests, leaving developers waiting 30 minutes or several hours before they can merge. AI adds more changes, more often, from more places — multiplying the number of validation jobs competing for resources.

Why are AI coding agents putting pressure on CI?

The increase in software activity is already visible. GitHub’s 2025 Octoverse recorded an average of 43.2 million merged pull requests each month—up 23% year over year—and almost 1 billion commits during the year, up 25.1%. GitHub notes that these are observational signals rather than proof that AI caused the growth, but either way, development volume is rising.

DORA’s 2025 research describes AI as an amplifier of an organization’s existing strengths and weaknesses. A fast, automated delivery system can absorb more output. A slow pipeline becomes more congested.

This changes CI from a developer-tooling concern into a capacity-planning problem. Multiple developers may run multiple agents at once, with every pull request triggering its own builds and tests. The result is highly concurrent demand that can arrive without much warning.

Why does CI need different infrastructure?

Production applications typically run continuously, maintain state, and prioritize availability. A CI job, on the other hand, is temporary: It starts, performs a defined task, and disappears.

That creates a very different demand pattern. CI workloads can jump from zero to thousands of vCPUs in seconds, run for a few minutes, then return to zero. Designing around average utilization misses the point; the infrastructure has to absorb the spike.

Compute is only one part of this path. Each job may need to pull a repository, download dependencies and container images, access cached build data, and upload artifacts. If the compute scales but the data can’t move quickly enough, the queue simply shifts to storage or the network.

Effective CI infrastructure therefore favors fast startup, workload isolation, high-performance compute, strong caching, and rapid access to data. The best architecture may not look like the one used for a persistent production application.

How can teams prepare CI for AI-generated code?

Teams can start by reviewing the full validation path:

• Measure the wait: Track queue time, runner startup time, test duration, cache performance, and artifact transfer time.
• Plan for concurrency: Model how the system behaves when many developers and agents submit work together, not only during an average hour.
• Keep data close: Place caches, dependencies, and artifacts where runners can access them quickly.
• Automate quality controls: Use unit tests, end-to-end tests, security checks, and risk-based review policies as guardrails for generated code.
• Match infrastructure to the workload: Assess whether the current compute, storage, and network design can scale quickly enough for short, intense bursts.

More code needs faster validation

AI coding agents could eventually work continuously, with agents coordinating other agents over longer tasks. Whether that arrives quickly or gradually, validation will remain the control point between generated code and dependable software.

The advantage won’t come from producing the most code. It will come from testing that code quickly, moving the required data efficiently, and enforcing quality without slowing delivery to a crawl.

For a deeper discussion, listen to our Uplink episode “CI Can’t Keep Up With AI”, where our CEO Michael Reid speaks with Aditya “JP” Jayaprakash, Co-founder and CEO of Blacksmith, about CI, AI coding agents, and the infrastructure behind modern software delivery.

Subscribe to Uplink wherever you get your podcasts.