Harness Agent-Ready SCM: What 7x More Pull Requests Broke in CI

Harness’s engineering organization opened about 7x more pull requests over the past year while the number of active contributors grew only about 30%. CI executions rose nearly 7x too, and at peak demand pipelines queued even though compute was available. For platform teams, Harness’s own numbers show where coding agents put pressure first: not on build capacity, but on concurrency, clones and review.

What Happened

In a Harness engineering post, the company shares data from running its own software factory, where its agents work alongside its Repositories and Build products:

Harness reads the read-heavy pattern as CI clones, automated tests and agents reading code rather than people uploading it. The peak queueing came from concurrency limits on a surge of parallel PR checks, and error rates stayed stable. Harness cleared the backlog by raising concurrency limits and temporarily disabling high-volume, non-critical triggers.

Why It Matters

Harness explains the mechanism: an agent runs the loop far more often than a person. It pushes small changes, opens a pull request, waits for checks, reads failures and pushes again, so one task can drive several pushes, CI runs and fresh clones. A single commit can start as many as 100 CI pipelines, and each may begin by cloning the repository.

That shaped Harness’s fixes for what it calls Agent Ready SCM, launched in August:

• Clone caching: a short-lived cache keyed on commit and client flags serves identical clone requests without rebuilding the packfile, improving cloning latency by 40%
• Cheaper merge checks: dry-run mergeability checks no longer take a distributed write lock, cutting merge-verification work by about two-thirds
• Merge Queues for busy repositories with concurrent merge requests
• CI scaling across regions and availability zones, with warm pools and a predictive algorithm that scales ahead of historical peaks

Harness also lets agents read repositories and pull request data without cloning, through its MCP server and CLI.

The Different Path From Anthropic

Harness compares its approach with Anthropic’s, which Stack Archive covered in Anthropic Scion. Anthropic built a custom REST API approach after merges jumped from roughly 2,500 to more than 30,000 a day. Harness argues that most teams can get much of that benefit by scaling Git horizontally instead of replacing it, though that is a vendor’s view of its own product. Harness also says review is the next bottleneck, and that its AI Code Review tool has increased internal merge speed by about 5x.

Source: Harness Blog

Content created with AI assistance and reviewed for accuracy.