Guiding Discovery through BixBench2 as if each problem was a mini-project

[Steven Truitt (stevetru1)]

Which Microsoft Discovery surface?

Microsoft Discovery app (Windows desktop client)

What did you build?

I used Discovery (App) to work through BixBench as if each question was a small project to tackle and create an evidence trail for. I wasn't particularly focused on maximum agentic performance and was more curious about how I could multi-thread small efforts using the engine while also ensuring there are critical feedback loops to spot issues. The overall findings were fairly interesting so I then asked to extend that work into a summary paper I thought might be interesting for everyone.

bixbench_discovery_report.pdf

How did you build it?

The approach here was very straightforward (and is described in more detail in the paper): take the benchmark and create a task for each question, then have the engine work through the full project through agent delegations. The interesting part was that I then duplicated the task tree for a follow-up critique and asked for consistency and attention to detail in how the final answers were generated (like a student taking a test).

Who might want to reuse this?

Hopefully this is useful for anyone looking at similar benchmarks or that has a project with multiple detailed and only semi-related questions to investigate.