v3 vs v4: Choosing a Computer Use Model on the API
You want to drive real desktops, browsers, and terminals without brittle selectors. The Coasty computer use API gives you two agent modes: v3 (guided) and v4 (autonomous with a pass/fail verifier). This post shows what each model does, when to use it, and how to call it with real requests and responses.
What v3 and v4 are
- ●v3 is a guided agent. You submit a task and optional instructions. The server drives the agent step by step, returning actions for you to execute on the remote machine. You control timing and can inspect progress via GET /v1/runs/{id}/events.
- ●v4 is an autonomous agent with a built-in verifier. You pass cua_version=v4. The server runs until the verifier succeeds or fails, then reports the final state. This is ideal for closed-loop tasks where you want a single endpoint call and a clear pass/fail outcome.
- ●Both models are billed at $0.05 per agent step on POST /v1/runs. The difference is control: v3 lets you pause and resume; v4 completes autonomously and reports success or failure from the verifier.
The key difference in the request
- ●Both models use POST /v1/runs with machine_id, task, and cua_version.
- ●v3: cua_version=v3 (or omit, since v3 is the default). You get a running task with intermediate events and can call POST /v1/runs/{id}/cancel or POST /v1/runs/{id}/resume.
- ●v4: cua_version=v4. The server runs a verifier against the final state and returns succeeded, failed, or another terminal state. No need to resume after completion.
- ●Optional fields like instructions, system_prompt, max_steps, deadline_seconds, and on_awaiting_human apply to both modes. on_awaiting_human can be pause, fail, or cancel when the agent needs human input.
# Example: launch a guided v3 agent on a cloud machine
# Replace YOUR_MACHINE_ID and YOUR_API_KEY below
COASTY_API_KEY="${COASTY_API_KEY:-$COASTY_API_KEY}"
# v3: guided agent (default)
curl -s https://coasty.ai/v1/runs \
-H "Authorization: Bearer $COASTY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"machine_id": "YOUR_MACHINE_ID",
"task": "Open Chrome and navigate to https://example.com",
"cua_version": "v3",
"max_steps": 50,
"on_awaiting_human": "pause"
}' | jq .# Example: launch an autonomous v4 agent with a verifier
# Replace YOUR_MACHINE_ID and YOUR_API_KEY below
curl -s https://coasty.ai/v1/runs \
-H "Authorization: Bearer $COASTY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"machine_id": "YOUR_MACHINE_ID",
"task": "Open Chrome, navigate to https://example.com, and click the Search button",
"cua_version": "v4",
"max_steps": 100,
"on_awaiting_human": "pause"
}' | jq .Check a v3 or v4 run
- ●GET /v1/runs returns a paginated list of runs.
- ●GET /v1/runs/{id} returns the status and metadata for a specific run.
- ●GET /v1/runs/{id}/events streams Server-Sent Events (SSE). Use Last-Event-ID header to reconnect. States include queued, running, awaiting_human, succeeded, failed, cancelled, timed_out.
- ●v4 reports final success or failure from its verifier. v3 reports completion once the task finishes; you can resume or cancel before final status.
v3 for guided control and resume; v4 for autonomous, verifiable workflows.
Where this beats brittle automation
Traditional automation relies on explicit selectors and fixed paths. If a UI changes or layout shifts, the script breaks. The Coasty computer use API drives agents that see the screen, interpret instructions, and adapt to changes. v3 lets you observe and decide when to resume, while v4 closes the loop with a verifier. This is powerful for UI-heavy workflows like forms, dashboards, and multi-step setups where reliability matters more than raw speed.
Pick v3 when you want guided control and the ability to pause or resume, and pick v4 for fully autonomous, verifiable tasks. Use POST /v1/runs with the right cua_version to get started. Get your API key at https://coasty.ai/developers and build robust browser, desktop, and terminal automation.