Database › Databricks
Databricks
No mandate reported. Databricks built an internal benchmark from its own merged PRs to choose which coding agents and models its engineers use day to day.[1]
Tools
- Coding agents benchmarked on an internal codebase (GLM 5.2 as a daily driver; Claude Opus)[1]
- coSTAR agent testing framework[2]
Metrics tracked
- On an internal benchmark built from merged PRs in its multi-million-line codebase, GLM 5.2 tied Claude Opus 4.8 at $1.28 versus $1.94 per task, and Databricks deployed it as a daily driver.[1]
Guardrails
- The coSTAR framework tests agents against scenarios before shipping, so changes can ship fast without breaking things.[2]
Hiring, reviews & process changes
Nothing verified from a public source yet. Know a source?
Reported results
- The agent harness mattered as much as the model: one harness (Pi) was about 2x cheaper for similar results.[1]
- Agent verification time fell from about two weeks to hours with coSTAR.[2]
Metric tags: Cost per PR, Quality / defects, Cycle time
Sources
- Databricks blog: Benchmarking coding agents on a multi-million-line codebase · 2026-07-08 · company source
- Databricks blog: coSTAR, how we ship AI agents at Databricks fast without breaking things · 2026-03-20 · company source
Last verified 2026-10-25. Industry and size are editorial bands; see methodology. Spot an error? Submit a correction.