jev| Part of the agent | Responsibility |
|---|---|
| Planning and writing | Existing agent |
| A defined route decision | Candidate Jev step |
| Permission to execute | Application code |
| Success measure | Completed user task, including recovery |
Find a decision you can name
Suppose an agent can search order records, explain policy or escalate a case. Jev could suggest which handler to use. Planning the conversation and writing the reply still belong to the surrounding agent. Start with one such decision, not a wholesale replacement.
Write down what the extra call should save
Perhaps your agent currently invokes a costly tool for messages that could be answered from a policy page. A routing step might avoid those calls. It might also choose the wrong handler and force a retry. Record both outcomes instead of counting only tokens saved on successful cases.
Compare completed tasks
Run the same tasks with the original agent and the version with routing. Keep its model, tools and permissions fixed. Count correct completions, unnecessary tool calls, recovery steps and total time. Inspect the trace when the versions disagree; a correct-looking route can still lead to an unsuccessful task.
Keep tool permissions in application code
A suggested handler is not permission to perform an action. Check the handler against an allowlist and enforce the user’s permissions before execution. If the decision fails, use a defined fallback to the original agent or review. The downloadable router demonstrates suggestions without executing tools.