Documentation review: 10 October 2026. Source: BC-Bench 0.15.0, commit
c84b793a8a9b7422af9512dfe67baeb2bc666d95. This update did not execute installation, agents, containers or benchmark runs.
Microsoft BC-Bench compares coding agents on AL tasks. Bug fixing, test generation, code review and natural-language implementation have different contracts and scorers; compare within a category.
The current upstream README highlights Copilot CLI, Claude Code and the dedicated BC PR Review runner. The latter uses BC-ALAgents and BCQuality for code review. This guide does not claim a direct BC-Bench runner for every ALDC host, including Codex.
| Component | Role |
|---|---|
| ALDC | Development context, specialist agents, skills and workflows. |
| BCQuality | BC knowledge and cited source review. |
| BC-Bench | Controlled experiments and category-specific evaluation. |
Combining them is an experiment hypothesis, not an established improvement. A generated patch is not a successful build, test run or human approval. Dataset sizes and available models belong to the selected source revision.
Sources: README, categories, experiments. Next: setup.