GitHits migrated from Blacksmith to Avrea and cut test feedback time 27%

Results snapshot

• Test feedback 27% faster at the median and 34% faster at p90, on a test suite that grew 22%
• Docker build times became predictable: 73% of builds finish in 1–2.5 minutes, up from 45%
• Runner pickup in 16 seconds (median)
• Production CI and release builds migrated within hours

‍

About GitHits

GitHits gives AI coding agents access to open source code and packages: real implementation examples, code search and exact line reads without cloning, package inspection (dependencies, vulnerabilities, changelogs, upgrade changes), and docs. It builds a version-aware index of a repository on demand, usually in 10-20 seconds. Works over MCP with Claude, Codex, Cursor, and the rest.

• Industry: Developer tools
• Stack: Elixir and Phoenix backend, Docker builds, GitHub Actions, ARM release builds

When agents write the code, CI sets the pace

GitHits is built for agent-driven development, and the team develops that way too: agents and humans both push code, and nothing merges without a green CI run.

"Fast CI pipelines are critical for the overall agentic development feedback loop," says Juha Litola, co-founder and Chief Architect of GitHits. "The faster humans and agents get feedback that the code works on a standard setup, the faster we are able to iterate and deploy."

GitHits was a happy Blacksmith customer, but recurring reliability issues (jobs queuing without runners and slow runs) triggered an evaluation of alternatives in early June 2026.

A head-to-head on the production repo

Juha ran Avrea against the existing runners on the production backend and compared job-for-job. In a step-level comparison of an identical job on the same vCPU class, Avrea removed about 15 seconds of runner platform overhead per job, with compute-step performance identical. The numbers made the decision easy.

Migration was a runner label change per workflow and was done quickly. Production CI moved over within hours of the first test run. Release builds went to Avrea's Linux ARM runners on Apple M5 Max the same day Juha tried them. The included 25GB of cache doesn't hurt either. On Blacksmith that was a paid add-on.

‍What changed after the switch

A month of production data on each side of the migration (successful runs only):

• Test feedback: 27% faster at the median (34% at p90) while the suite grew from ~9,700 to ~11,800 tests
• Docker builds: the middle half tightened from a 1.4–2.9 min spread to 1.8–2.0 min
• Pickup: 16 seconds median from queued to runner start; 98% of ~1,200 jobs within 2 minutes

Part of the test feedback gain is shared credit: the precommit pipeline was restructured into parallel stages during the migration, a split that Avrea's cheap per-job cache restores made viable.

Pickup matters just as much. A job stuck in a queue leaves an agent hanging mid-task, just like it does a human. Since the switch, the median wait for a machine is 16 seconds.

"Avrea's pricing is similar to Blacksmith's, but the runners are faster and the Docker layer caching is more effective. We've also had some queuing with Blacksmith runners where jobs take minutes to start, but with Avrea things have been a breeze so far."

Juha Litola, Co-founder and Chief Architect, GitHits

Run the same head-to-head on your own repo. Benchmark Avrea on one workflow and compare the numbers.

‍

‍