[NickGuAI]

Students build a deep research agent with at least five research sub-agents searching in parallel, then compare it with one agent on three fixed DeepResearchBench tasks. Any architecture or framework is allowed.

The assignment contains the complete original prompts for tasks 99 (Fashion & Beauty), 93 (Software), and 100 (Social Life), plus machine-readable task records. The evaluation SOP gives setup and execution commands for the pinned, unchanged RACE evaluator. The helper packages six reports, checks that all six scores and four dimensions exist, and produces an optional comparison table.

Graded item	Points
Actually run at least five research sub-agents searching in parallel	8
Run single and multi on all three tasks, score the six reports with RACE, and compare results	2

One report per task and mode is required; repeats are optional. There is no score threshold, class ranking, extra checkpoint, or prescribed architecture. Students complete EXPLANATION.md: what they built, why they chose the roles and architecture, and the single/multi comparison. README holds setup and run commands. They submit through their team's private repository.

Files

• studio/README.md: navigation entry.
• studio/studio-06/instruction/: assignment, two-row rubric, full prompts, task JSONL, evaluation SOP, helper and tests, and upstream license.
• studio/studio-06/submission/: submission guide, the required EXPLANATION template, run instructions, and an optional evidence example.
• studio/studio-06/README.md and .gitignore: overview and local-work exclusions.

Validation

• All three complete task records match upstream revision f2735b5c3636759c22f9e936c0232de8bf545b67, with matching original criteria and references.
• Six unit tests pass, including prompt/report preservation, missing scores/dimensions, and native model settings.
• Offline integration exercised the helper through the original cleaner, judge wrapper, weighting, native result files, and comparison writer with 12 canned HTTP replies and network access prohibited. All six fixture scores were 0.5 as expected.
• Relative links resolve; git diff --check passes; reviewed files contain no credentials. The EXPLANATION questions match the main instruction verbatim, and all repository deadline references agree.
• This validates the evaluation wiring. It is not a paid end-to-end benchmark or a performance result. Publication does not depend on the separate instructor model experiment.

Publication state

The instructions, overview, and submission guide set the deadline to Friday, October 23, 2026, 13:10 ET (17:10 UTC), before class, following the existing Studio convention and instructor direction. The live Lecture 06 Studio slides and these instructions both use 8 + 2 = 10 points. The live slides do not list a deadline; CourseWorks has no Studio 06 assignment yet. That assignment must use the same deadline and 10 points. This PR does not create it, change the deck, merge, or publish private grading criteria.

The older assignment checker assumes a checkpoint-plus-explanation rubric. It was not changed or represented as passing this owner-requested 8+2 design.