Vercel CI beta targets agent-heavy repos with 2-6x faster push-to-pass times

Vercel has launched Vercel CI, a continuous integration service now in beta that is explicitly pitched at repositories where coding agents, not only humans, are pushing most of the code. According to the Vercel blog, the service runs only the tasks a change actually affects and allocates only as much compute as those checks require. On Vercel's own repositories, the company says the time from git push to a passing CI status is now two to six times shorter than it was previously.

The framing in the October 28 announcement is blunt: AI coding agents now submit a rapidly growing share of pull requests, and every one of them waits on CI before a human can review or merge it. Most CI systems re-run an entire workflow on every push, even when the change is a single line, so the more often agents push, the more time a team spends waiting in line.

How it works

Vercel CI is built on Turborepo and reads turbo.json to model a repository's tasks and the dependencies between them. Existing Turborepo users get CI with no separate configuration file; everyone else adds a turbo.json file and the turbo package, which Vercel notes also works in single-package repositories.

Instead of maintaining workflow YAML, developers tag the tasks they want to run in CI — for instance, marking a lint task with a ci tag. When code is pushed, Vercel CI works out which tasks the change affects. Anything whose inputs are unchanged is pulled from the cache rather than re-executed, with Vercel's Remote Cache switched on by default. The remaining tasks each run on their own machine on Hive, Vercel's compute layer, starting as soon as their upstream dependencies finish. Vercel's changelog adds that the service installs dependencies once per run and caches them across subsequent code updates.

Observability and GitHub integration

Once configured, runs can be inspected through the Vercel dashboard, the Vercel CLI, or Vercel's MCP server — meaning agents themselves can query CI state and logs. The new CLI surface includes commands to list runs, inspect a run's tasks, fetch logs with per-task and log-level filters, and search across recent run logs, which Vercel positions as a way to check whether a given test has failed repeatedly across the week.

On the collaboration side, Vercel CI posts a check to GitHub pull requests, so existing branch protection rules keep blocking merges when CI fails, unchanged from how teams work today.

Scale and what's next

Vercel says the service has already completed nearly 8 million tasks across its own repositories and the private beta that preceded today's launch. Heading toward general availability, the company plans advanced observability, a new vc ci CLI command for distributed execution before code is pushed, and more granular test selection. The beta is available on all Vercel plans.

Why it matters

Vercel's core argument is that in agent-heavy development, CI has become the rate limiter. An agent that pushes code and then idles while a full workflow re-runs completes far fewer useful iterations per hour than one that gets near-instant feedback, and human reviewers sit in the same queue. Affected-task execution combined with aggressive caching is the established remedy — Turborepo users already benefit from it locally — but Vercel is packaging it as a hosted, zero-YAML CI with an interface that agents can query programmatically via MCP, closing the loop from push to diagnosis without a human in between.

Two caveats are worth holding onto. The 2-6x figure is Vercel's own measurement on its own repositories, not an independent benchmark, and the product is in beta with features still to come. Even so, the launch is a clear signal of where CI tooling is heading: designed for machines that push constantly, read logs themselves, and never tire of waiting — but still burn time and money every time they do.