Pick the model, tools and skill for each agent request
Before an agent runs, one Jev call decides which model tier, tools and skill the request needs, so simple asks stay cheap.
Author's figure: Author reports, on their own Claude Code sessions, the right skill picked 90.0% of the time versus 59.4% for the baseline, with under-provisioned turns down from 19.9% to 3.3%.. Reported by the builder, not measured by us.
The use case
Teams building AI agents with dozens of tools send every tool and the most expensive model to every request. One Jev call reads the incoming request and answers which tools apply, which skill fits, and how hard the turn is; code turns that into a model tier and effort budget. A video-editor team reports using Jev to pick a subset of 100+ tools and to tell small talk from real work. A regex fallback runs if the call misses its deadline. Engineers review misrouted turns from logs and adjust the catalog.
Example workflow
- Define a catalog of models, tools and skills with short descriptions
- On each request, ask Jev tool, skill and difficulty questions in one call
- Code maps answers to a model tier, effort budget and tool list
- Run the agent turn with only the chosen tools
- Review misrouted turns in logs and update catalog descriptions
Tools
- Jev (TypeSafe AI)
- Claude Agent SDK
- jev-harness-router (npm)
Risks and guardrails
- Under-provisioning a hard request gives poor answers
- Missing a needed tool stalls the agent
- Catalog descriptions need upkeep as tools change
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- GitHub repo jev-harness-router by JoacoMarc: github.com/JoacoMarc/jev-harness-router
- Hacker News comment by dthedavid on choosing tool subsets for a video editor: news.ycombinator.com/item