Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Multi-agent system research agent for CoWork OS. Use when: researching multi-agent papers, frameworks, production case studies; maintaining a research queue; producing CoWork OS applicability analysis and implementation recommendations. Triggers: 'multi-agent research', 'research multi-agent systems', 'multi-agent papers', 'agent orchestration research', 'CoWork OS research'.
.claude/skills/cowork-os-cowork-multi-agent-research/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -37% | 0% |
You are the CoWork OS Multi-Agent System research agent. Follow these instructions to conduct systematic research on multi-agent systems and produce actionable findings for CoWork OS.
PROGRESS.md before any research. Never re-research a topic already marked done.research/YYYY-MM-DD-topic-slug.md.| File | Purpose | |------|---------| | PROGRESS.md | Research queue, completed topics, next priorities | | research/YYYY-MM-DD-topic-slug.md | Individual research documents |
Default paths are relative to the workspace root (project root or ~/.cowork/workspace). If PROGRESS.md does not exist, create it with the template from references/full-guidance.md.
research/, PROGRESS.md updated, summary of findings and next topics.Each research document must include:
For detailed workflow, PROGRESS.md template, and research document template, see references/full-guidance.md.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-10 | fail→pass | 9,064 | 3,619 | -60% | 1 | 1 | 0% | 1,572 | 1,149 | -27% | 0 | 0 | — |
case-11 | fail→pass | 7,458 | 6,683 | -10% | 1 | 1 | 0% | 1,269 | 1,740 | +37% | 0 | 0 | — |
case-20 | pass→pass | 12,574 | 9,152 | -27% | 1 | 1 | 0% | 1,970 | 2,179 | +11% | 0 | 0 | — |
case-21 | pass→pass | 16,002 | 12,603 | -21% | 1 | 1 | 0% | 2,698 | 2,735 | +1% | 0 | 0 | — |
case-01 | fail→fail | 5,728 | 5,688 | -1% | 1 | 1 | 0% | 248 | 848 | +242% | 0 | 0 | — |
case-02 | fail→fail | 4,999 | 5,662 | +13% | 1 | 1 | 0% | 193 | 857 | +344% | 0 | 0 | — |
case-03 | fail→fail | 31,393 | 4,778 | -85% | 1 | 1 | 0% | 5,954 | 897 | -85% | 0 | 0 | — |
case-04 | fail→pass | 7,266 | 3,796 | -48% | 1 | 1 | 0% | 1,117 | 1,206 | +8% | 0 | 0 | — |
case-09 | fail→pass | 10,458 | 6,922 | -34% | 1 | 1 | 0% | 1,576 | 1,702 | +8% | 0 | 0 | — |
case-05 | fail→pass | 8,028 | 2,267 | -72% | 1 | 1 | 0% | 1,457 | 918 | -37% | 0 | 0 | — |
case-06 | fail→pass | 6,944 | 4,148 | -40% | 1 | 1 | 0% | 1,165 | 1,338 | +15% | 0 | 0 | — |
case-07 | pass→pass | 9,379 | 3,892 | -59% | 1 | 1 | 0% | 1,556 | 1,261 | -19% | 0 | 0 | — |
case-08 | pass→fail | 11,062 | 5,790 | -48% | 1 | 1 | 0% | 1,719 | 1,472 | -14% | 0 | 0 | — |
case-12 | pass→pass | 9,126 | 6,038 | -34% | 1 | 1 | 0% | 1,370 | 1,610 | +18% | 0 | 0 | — |
case-13 | fail→pass | 10,455 | 8,746 | -16% | 1 | 1 | 0% | 1,577 | 1,933 | +23% | 0 | 0 | — |
case-14 | pass→pass | 15,300 | 8,619 | -44% | 1 | 1 | 0% | 2,425 | 2,086 | -14% | 0 | 0 | — |
case-15 | pass→fail | 9,679 | 4,420 | -54% | 1 | 1 | 0% | 1,414 | 1,224 | -13% | 0 | 0 | — |
case-16 | fail→pass | 6,797 | 1,730 | -75% | 1 | 1 | 0% | 1,080 | 839 | -22% | 0 | 0 | — |
case-17 | fail→pass | 13,048 | 3,188 | -76% | 1 | 1 | 0% | 2,179 | 1,071 | -51% | 0 | 0 | — |
case-18 | pass→pass | 6,980 | 2,357 | -66% | 1 | 1 | 0% | 918 | 1,008 | +10% | 0 | 0 | — |
case-19 | fail→pass | 8,584 | 6,133 | -29% | 1 | 1 | 0% | 1,522 | 1,533 | +1% | 0 | 0 | — |
case-22 | pass→pass | 8,752 | 6,018 | -31% | 1 | 1 | 0% | 1,555 | 1,788 | +15% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 19 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.