Skip to content

Add fix-bug custom agent experiment arm - #845

Open
ventselartur wants to merge 8 commits into
mainfrom
private/ventselartur/BugFixingSkillToBCBench
Open

Add fix-bug custom agent experiment arm#845
ventselartur wants to merge 8 commits into
mainfrom
private/ventselartur/BugFixingSkillToBCBench

Conversation

@ventselartur

@ventselartur ventselartur commented Sep 2, 2026

Copy link
Copy Markdown
Collaborator

Why

The bug-fix experiment should explicitly select a dedicated workflow without adding a new per-skill selection mechanism. BC-Bench already supports custom-agent experiment arms through agents.enabled and agents.name, which copy the profile agents and invoke the harness with --agent=<name>.

This adds a standalone fix-bug custom agent containing the trimmed offline workflow from microsoft/BCAppsBugFix: investigate and plan, implement in a self-correcting loop, and validate with AL tools when available. Work-item fetch, test baseline, Miapp propagation, critique, AL review, and pull-request phases remain excluded so the agent fits BC-Bench's unattended evaluation model.

Summary

  • Added the modular fix-bug custom agent to the microsoft-BCApps and microsoftInternal-NAV profiles
  • Configured the experiment with skills.enabled: false, agents.enabled: true, and agents.name: fix-bug
  • Kept the standard bug-fix prompt unchanged so workflow selection comes from --agent=fix-bug
  • Removed the out-of-scope skills.names selection and skill-name result metadata changes
  • Removed premature category documentation pending experimental results
  • Added tests for custom-agent configuration, prompt behavior, and copying the agent's supporting files

Validation

  • uv run pytest -q
  • uv run ruff check .

@haoranpb Sun Haoran (haoranpb) left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great that you added the skills!

As discussed yesterday, let's keep this simple so we can quickly iterate.

Small tip, we could trigger runs with Claude Code where MCP is not blocked

Comment thread EXPERIMENT.md Outdated
Comment thread src/bcbench/agent/shared/config.yaml Outdated
Comment thread docs/bug-fix.md Outdated
@haoranpb

Copy link
Copy Markdown
Collaborator

ventselartur The Claude Code run finished: https://github.com/microsoft/BC-Bench/actions/runs/33734248178

I will be working on getting more information of what happened during the run (i.e. the transcript)

You can take a look at the results in the meantime

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@ventselartur ventselartur changed the title Add bc-fix-bug skill arm and per-skill selection for experiments Add fix-bug custom agent experiment arm Sep 3, 2026
ventselartur and others added 3 commits September 3, 2026 14:14
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants