Many support, ops, and sales teams still rely on web interfaces. To make LLM agents truly useful, they must interact directly through a browser.

We build a browser‑using agent864 with OpenAI Agents SDK and Playwright MCP. The agent runs in a loop: task + browser state → interpret → action → new state until the task is finished. Two links are required: an observation channel (e.g., structured DOM) and an action channel (element‑targeted commands). We focus on structured page observations paired with element‑targeted actions.

Case study: resolve a support request in a local console served at http://127.0.0.1:8000. The agent’s task is to open the app, address case ORD-1042, add an internal note, and confirm resolution. The agent uses GPT adjustable‑5.4 with minimal instructions “You are an agent that can interact with a web browser.” and a Playwright MCP server exposing Chrome tools.

Running the agent launches Chrome, performs clicks and typing automatically, and logs each step. The loop continues until the task description signals completion, demonstrating how to give an LLM agent direct browser control using the SDK and MCP.