Vercel CI enters beta: runs only the Turborepo tasks a change affects, 2-6x faster CI built for coding agents

Background and pain point

Coding agents are opening more pull requests than ever, and each one waits on CI before it can be reviewed or merged. Most CI systems rerun the whole workflow on every push, even for a one-line change. Vercel's post puts it plainly: the more your agents push, the longer people wait on CI, so CI is becoming the bottleneck in agent-driven development.

Architecture and mechanism

Vercel CI uses Turborepo to understand tasks and their dependencies. It reads turbo.json directly, so there is no separate CI config, and it works in single-package repos too. On each push it works out which tasks are affected; tasks whose inputs have not changed come from Remote Cache, and the rest each run on their own machine on Hive, starting as soon as their dependencies finish and getting only the compute they need, much like Fluid compute scales with traffic.

Benchmarks

Vendor figures only: push-to-"CI passed" time in Vercel's own repos is 2-6x shorter, and nearly 8 million tasks have run across those repos and the private beta. There is no third-party data or like-for-like comparison with GitHub Actions or similar tools yet.

Getting started

Add "tags": ["ci"] to the tasks you want in turbo.json, or have your coding agent run npx skills use vercel/ci@setup in the repo and follow the skill to pick tasks, configure CI and open a pull request. Then use vc ci ls, vc ci inspect, vc ci logs and vc ci grep to inspect runs and logs, or hand debugging to an agent through Vercel MCP. Results post back to GitHub pull requests as a check. Before general availability Vercel plans deeper observability, distributed execution before you push, and more granular test selection.