LLM API for browser agents

Browser agents in a production application

A procurement assistant opens an approved supplier portal, finds a specific part, fills a purchase request, and stops at the final Submit button. Browser state is exposed through tools such as open_page, inspect_page, click, and fill.

Sol plans the long interaction, but the browser controller—not the model—enforces domain allowlists, blocks file downloads, records screenshots, and requires confirmation immediately before submission.

How the browser agents workflow operates

• Create an allowlisted browser session with a recoverable test account.
• Expose semantic page state and a small action set.
• Execute one action, then inspect the new state.
• Recover from layout or timing changes without guessing coordinates.
• Stop immediately before the consequential action and request confirmation.

Call gpt-5.6-sol through LLMFly AI

Send the request from your server. Replace the example content and placeholder tool schema with data and tools from your application.

curl https://app.llmfly.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMFLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [
      {"role": "system", "content": "Complete the procurement form with browser tools. Stay on allowed domains. Stop before the final submission and ask for confirmation with a summary of the exact order."},
      {"role": "user", "content": "Prepare a purchase request for the approved part and quantity in the attached specification."}
    ],
    "tools": [{"type": "function", "function": {"name": "YOUR_TOOL", "description": "Replace with the tool contract for this workflow", "parameters": {"type": "object", "properties": {}, "additionalProperties": false}}}]
  }'

Why gpt-5.6-sol is the primary model

Sol is used for long, stateful browser work where recovery and stopping at the correct state matter more than raw token cost.

Terra can handle short, stable navigation flows. Do not downgrade tasks that submit, purchase, delete, or send without rerunning the safety evaluation.

Acceptance checks for browser agents

Failures to handle before deployment

• Using production accounts for evaluation
• Relying on coordinates alone
• Confirming too early
• Skipping final-state verification

Returned output and run records

A verified filled form and final-state summary, with no submission before confirmation.

Record the model ID, request ID, token usage, retries, validation result, and final disposition for every production run.