When you call POST /v1/runs you choose a CUA version (cua_version). v3 is a standard computer use model focused on reliable, guided task execution. v4 adds an autonomous pass/fail verifier and can run without your intervention. The difference shows up in how the server drives the agent, how it handles delays, and the billing per step. This post explains the exact request fields, pricing, and when to pick each.
Request fields for CUA version
- Base URL: https://coasty.ai/v1
- Authorization header: Bearer <key> or X-API-Key: <key>
- POST /v1/runs body fields: machine_id, task, cua_version (string, default "v3", can be "v4"), instructions (optional string), system_prompt (optional string), max_steps (optional integer), deadline_seconds (optional integer), on_awaiting_human (optional string: "pause"|"fail"|"cancel"), webhook_url (optional string)
- v4 enables autonomous mode with a pass/fail verifier that waits for a final status before marking the run as succeeded or failed
- Both versions bill $0.05 per agent step, but v4 may complete more steps autonomously reducing manual oversight
curl -X POST https://coasty.ai/v1/runs \
-H "Authorization: Bearer $COASTY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"machine_id": "aws-ec2-eu-west-1",
"task": "Open Chrome and navigate to https://example.com",
"cua_version": "v4",
"instructions": "Do not close the browser after completing the task",
"max_steps": 100,
"on_awaiting_human": "pause",
"webhook_url": "https://your-app.com/callbacks/coasty"
}'Behavioral differences
- v3: The server follows your instructions and task description step-by-step. It often shows intermediate states and can stop on awaiting_human events unless you set on_awaiting_human to fail or cancel.
- v4: Runs autonomously with a built-in pass/fail verifier. The server decides when the task is done and returns a final status without your manual intervention. It is ideal for production workflows where you want a completed run with a clear success/fail outcome.
- Both versions stream events via GET /v1/runs/{id}/events (Server-Sent Events) so you can monitor progress in real time.
Use v3 for predictable, guided tasks with human checkpoints; use v4 for autonomous, production-ready runs with an automatic verifier.
Where this beats brittle automation
Traditional automation relies on selectors, IDs, or hardcoded API calls. Those break when UI changes or when the app does not expose a programmatic interface. A computer use model like v3 or v4 reads the screen, understands natural language instructions, and clicks, types, and navigates like a human. You can ask it to 'open Chrome, go to https://example.com and click the first link' without knowing the link’s exact selector. This lets you build agents that adapt to layout changes and handle complex workflows with fewer brittle rules.
Pick v3 for guided, controllable tasks and v4 for fully autonomous runs with automatic success/fail verification. Both share the same $0.05 per step billing. To get started, create a key at https://coasty.ai/developers and then POST /v1/runs with your chosen cua_version.
Want to see this in action?
View Case Studies