Finding the real, high-value uses of autonomous agents. We rank candidate use cases by value × demand × feasibility, pre-wire each with the skills it needs, and attach a runnable eval — capability verified against public benchmarks, not asserted.
A pre-validated starter set across 16 categories, mined from GitHub as the live index. Three unit types compose a managed agent — kept typed, not blurred:
Instructions — how to do a task. Progressive disclosure, ~0 context until triggered. Exposed as toggleable customization.
A callable capability — how the agent reaches a system. Background infrastructure, auto-mounted, never a catalog item.
Goal + skills + tools + loop. The primary gallery item. Cross-domain work delegates to peers via A2A.
GitHub is the index; awesome-lists are overlays. A scheduled discover → parse → rank pipeline scores repos by
log10(Stars+1) · e−λ·t · (1 + Commits₉₀ₐ/50) · Health
λ = 0.005 (≈138-day staleness half-life) · ×0.5 no license · ×1.2 CI + clean scan
Breadth isn't safe to expose. Of ~20k listings on the big directories, >60% are unmaintained forks or clones; a 1,899-server audit found 7.2% with vulnerabilities and 5.5% with tool-poisoning / prompt-injection.
Gate: static-verify + signature + container/WASM sandbox + approval on state-modifying actions.