Give the goal. Jev answers the hard question. Your code picks the call.
Jev is TypeSafe's decision model. It returns labels and confidence, never text. Your agent states a goal. Jev answers the judgment call. Vaaya pays and settles.
Scripted playback: prices are the catalog's per-call quotes as of September 23, 2026, and the decision and result lines are a worked example, not a recorded run
The sequence
Every run above follows the same five moves. The ambiguity goes to a model; the money never does.
- The agent proposes a call
It names the service, the action, the params and the quote it wants to spend against the goal.
- Code rejects the obvious
Input, permission, cache and budget checks run first. Duplicates, exhausted budgets and disallowed actions never reach a model.
- Jev judges the ambiguity
One state and up to eight typed questions per call, the contract of Vaaya's decisions action. Back come labels, probabilities and an escape option, never prose.
- Code applies the policy
A label is evidence for a branch, not authorization. Permissions and budgets stay in your application.
- Vaaya executes and settles
The permitted spend is reserved, the call runs against your Vaaya credit, and the final charge settles with a receipt.
Watch Vaaya work
The product film: agents running paid jobs through Vaaya inside claude.ai, VS Code and Claude Code. The gate above sits in front of calls like these.
Behind the demo
The demo compresses three published pieces. They carry the methods, the limits and the numbers.
- Jev model routing: what 450 real calls revealedA live 50-prompt benchmark: 34.7% lower usage cost, a 34% slower p95, and every call's receipt.
- Ask Jev before your agent spends a dollarThe spending gate the third demo plays: checks first, typed judgment second, the purchase behind ordinary code.
- Only pay to enrich the CRM records worth callingThe enrichment version: gate the record before paying any finder, then verify before anyone sends.