Overview & Result
As agent skill libraries grow to dozens or hundreds of capabilities, stuffing all schemas into the main LLM context causes severe token bloat, latency, and confusion. This router uses Jev Choice to pick the exact right skill in tens of milliseconds, loading only the necessary tool into context.
How Jev fits in the loop
- Maintain a lightweight index of skill slugs and short descriptions.
- Send incoming user task to Jev Choice with the available skill options.
- Obtain the winning skill with probability metrics.
- Dynamically inject the selected skill definition into the agent execution context.
How to reproduce
- Clone GodsBoy/jev-agent-skill-router and register multiple skill schemas.
- Send diverse user intents to the router and inspect the returned skill slug and confidence score.
- Verify that only the selected skill's definition is loaded into the agent's working context.
Why this build matters
Enables multi-skill autonomous agents to scale to hundreds of tools without suffering context window exhaustion or degradation in tool-calling accuracy.
Limitations
- Skills must have clear, non-overlapping boundary descriptions.
- Composite queries requiring chained skills must be decomposed before dispatch.
- Fallback logic is required when no skill matches with sufficient confidence.