Decisions and controlled autonomy
Open Decisions to connect Jev from TypeSafe or an alternative provider. Organisation admins save the key, choose a model, and set rollout policy. Model usage is billed by your provider; Lens does not include unlimited model credits.
Start with Observe
Observe records assessments while existing AgentShield enforcement stays in control. Automatic audit assessment is opt-in. Assist holds proposed actions for review. Limited Autopilot can approve explicitly selected exact scopes after reviewed evidence qualifies. Admins expand scopes themselves; the system never promotes itself into unrestricted autonomy.
Existing AgentShield blocks remain final. Baseline human-approval requirements stay human decisions. Provider failure holds actions in Assist and Limited Autopilot. The gateway must claim an exact, unexpired approval once before execution. Policy changes and pausing invalidate pending permissions.
Review and teach
Inspect the proposed action, context, recommendation, and disposition in decision history. Record a correction with an explanation. You can choose to reuse that correction as an example for the same exact scope. Later corrections preserve earlier reviews, can withdraw an unused approval, and can withdraw example consent. They cannot execute a completed action again.
Learning means retrieving reviewed examples; it does not retrain Jev's model weights. Each eligible scope needs twenty reviewed cases with safe reported outcomes and no unsafe approvals for the current provider and model. This is a pilot threshold, not a statistical safety guarantee. Outcomes must come from a trusted integration with independent application checks.
Explore the controlled demo
At app.aiagentlens.com/demo, open Decisions. Six fixed synthetic scenarios were assessed locally with jev-1.13.0, and their recorded answers are replayed in your isolated workspace:
| Scenario | Recorded Jev recommendation |
|---|---|
| Public README read | Allow |
| Production release without approval | Needs human approval |
| Fictional customer-credential export | Block |
| Destructive workspace deletion | Block (also blocked by deterministic policy) |
| Read-only synthetic inventory count | Allow |
| Installer command from an injected README | Block |
These are real recorded model responses to fictional inputs, not live visitor API requests or accuracy benchmarks. Try a correction; it changes only your temporary database. Provider keys, arbitrary assessments, execution claims, and external integrations are unavailable in the demo. No tools execute. Reset starts a clean copy. Every workspace is destroyed after one hour, even while active; a cleanup sweep runs every minute.
Plans
Team is USD $79 per named human user/month, or $59 per user/month billed annually, with a five-user minimum. Team access is arranged through [email protected]; there is no payment checkout yet. Enterprise fleet capacity is contracted separately. Free supports evaluation and Observe; enforcement requires a paid plan. API keys and service tokens are not human seats.
Before using production autopilot, validate unseen and adversarial cases, verify outcomes independently, and test expected fleet load. This initial version uses exact scopes and SQLite; it does not provide generalised permissions, automatic policy optimisation, or a production-capacity guarantee.