Natural language authoring without losing control

Natural language authoring is the feature every AI testing vendor leads with. Write a journey in English; get executable steps. The demo is impressive. The hard part is what happens while the draft is being built — and who still owns the save.

Without a human in the loop, AI-authored cases become a new kind of flaky debt: plausible, unreviewed, and silently wrong.

Sonar is an in-app assistant, not a silent autopilot#

Sonar is Orcastrator’s project-aware AI assistant. Open it from a project to:

• Chat about suites and cases that already exist
• Author or extend test cases with you watching
• Drive Live Mode so each candidate step is grounded in the real page

Sonar reasons and drives the authoring experience. Orca agents still execute the browser work. That split matters: AI proposes; the runtime proves; you decide what sticks.

Journey A: author while you watch the page#

Inside Orcastrator, keep the Test Case Editor open. Sonar proposes environment and agent, enables Live Mode, and adds ghost steps — candidates you accept, edit, or reject after they run against the live app.

Persistence stays a deliberate human action. You save when you are happy. Nothing becomes a published version until Test Management’s draft → review/publish loop says so.

Journey B: same substrate from your IDE#

Prefer Cursor or VS Code? Connect over MCP. An IDE agent can validate-then-mint against Orca without keeping an editor tab focused — useful when you live in the IDE and still want governed cases back in the project.

Use both: Sonar for in-browser collaboration, MCP when the editor is home.

Credits, models, and the publish gate#

Sonar draws on plan credits and optional model tiers. When credits run out, authoring pauses until the period resets or the plan changes. That is intentional: unlimited silent generation is how fiction ships.

Before you publish an AI-touched case, ask:

• Does it map to a Blueprint area?
• Are assertions business-meaningful?
• Did a human review the risky steps?
• Will the next orchestration run a published version Pulse can score?

Speed with a control plane#

AI should compress authoring time. The quality control plane should compress uncertainty. Natural language is the interface; live accept/reject and versioned publish are the guarantee.

That is how teams get the velocity of AI without shipping fiction.