v3 vs v4: Choosing a Computer Use Model on the API
When building a computer use agent, the choice between v3 and v4 changes your workflow from manual control to fully autonomous execution. v3 is for explicit human-like actions, v4 adds a pass/fail verifier and runs without human intervention. Both use the same POST /v1/runs endpoint, but the cua_version field and the verifier behavior differ.
How it works
You submit a task run to POST /v1/runs with a machine_id, the task description, and cua_version set to either "v3" or "v4". The server returns a run_id and begins execution. For v4 you can also supply system_prompt, max_steps, deadline_seconds, on_awaiting_human, and a webhook_url. The agent emits events via GET /v1/runs/{id}/events and bills $0.05 per agent step. When cua_version is "v4" the server includes a pass/fail verifier and does not stop for human approval unless you set on_awaiting_human to "pause".
curl -X POST https://coasty.ai/v1/runs \
-H "X-API-Key: $COASTY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"machine_id": "m-12345",
"task": "Open Chrome and navigate to https://example.com",
"cua_version": "v4",
"system_prompt": "You are an autonomous agent that drives the desktop.",
"max_steps": 50,
"deadline_seconds": 300,
"on_awaiting_human": "pause",
"webhook_url": "https://your-server.coasty-webhook"
}'
v3: manual control
- ●cua_version: "v3"
- ●Agent waits for each action to be sent or for human approval if on_awaiting_human is set
- ●Best for debugging, interactive workflows, and cases where you want to inspect or override every step
- ●Bills $0.05 per agent step, same as v4
v4: autonomous with verifier
- ●cua_version: "v4"
- ●Includes a pass/fail verifier that can stop the run based on success or failure criteria
- ●Does not pause for human approval unless you set on_awaiting_human to "pause"
- ●Ideal for long-running, unattended workflows and CI/CD pipelines
- ●Bills $0.05 per agent step, same as v3
For fully autonomous tasks, set cua_version to "v4" and optionally include on_awaiting_human: "pause" for safety.
Where this beats brittle automation
Computer use agents see the screen and act like a human, so they adapt to UI changes, dynamic layouts, and missing selectors. v4 adds a verifier that evaluates the final result, reducing false positives from brittle assertions. You can also drive real desktops and browsers via POST /v1/machines and use MCP to control Coasty from Cursor or Claude Desktop.
Start with v3 to control the agent step by step, then switch to v4 for fully autonomous runs with verification. Set cua_version appropriately and monitor events via GET /v1/runs/{id}/events. Get a key at https://coasty.ai/developers and start building reliable computer use agents.