▸case-12 For compliance and security auditing, we need to log every action performed by our desktop control agent. What specific data and artifacts should be captured for each executed action in an agent session log? | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-11 We are deploying a desktop automation agent across our enterprise and want to prevent unexpected API bill spikes caused by runaway agent sessions. How can we track and cap API expenses during agent execution in Python? | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-09 Our Anthropic computer use agent frequently fails when trying to click and select options in HTML select dropdown menus using pixel clicks. What reliable alternative strategy should be used to interact with dropdown menus and scrollable elements? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-05 My vision-based desktop agent executes actions rapidly, firing mouse clicks and key presses back-to-back immediately after receiving LLM responses. However, target applications often miss input or fail to update UI state in time. How should action execution timing be handled between steps? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-08 Our browser automation agent reads live web pages to execute user commands. Recently, an agent visiting a customer forum was tricked into navigating to a malicious URL and attempting file downloads because of text on the page. What architectural defenses should we implement against prompt injection embedded in web pages? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-10 Our desktop automation agent runs multi-step tasks requiring 40+ actions. As the task progresses, API costs skyrocket and the model begins repeating actions or losing track of context. How should image and history context be managed in long-running vision agent loops? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-07 We need to automate filling out a 20-field web application form using Claude Sonnet. Should we use full-screen vision screenshots and pixel coordinates or accessibility DOM snapshots with structured selectors, and how should this be implemented in Python? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-06 I built a visual AI agent that automates web interactions by calculating target bounding box centers and clicking exact pixel coordinates like (500, 300). Bot detection systems on target sites are flagging and blocking the agent. How should mouse click coordinate generation be modified to bypass coordinate-based bot detection? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-02 I'm building an AI agent that controls my desktop by capturing screenshots and issuing keyboard and mouse commands using Anthropic's API. I need a python script that implements the iterative screenshot-action loop, but includes an approval gate to pause and prompt the user whenever the model attempts sensitive actions like submitting forms or entering credentials. Please provide a python class for this agent setup with the human-in-the-loop validation logic. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-17 In Anthropic's computer use paradigm, when an agent needs to update source code or configuration files, should it type text keystroke-by-keystroke into a visual text editor application, or is there a dedicated tool pattern? How should file viewing and modifications be handled? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-20 Write an asynchronous Python script using Playwright to scrape product titles and prices from an e-commerce catalog page at https://example.com/products and save them to a JSON file. Do not include any LLM or AI decision-making. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-21 Create a FastAPI endpoint /api/v1/users that receives a JSON payload with user details, validates the input using Pydantic, and saves the record into a PostgreSQL database using SQLAlchemy. Include input validation. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-14 Our Anthropic computer use agent setup includes a bash tool implementation. How can we prevent the model from executing dangerous terminal commands that could corrupt the filesystem or system devices? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-22 Write a production Dockerfile for a standard FastAPI web application running on Uvicorn with Gunicorn. Use a multi-stage build, install requirements, and set the entrypoint. | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-03 I am writing a Python script for an AI vision agent that automates GUI interactions by taking 4K screen captures and sending them to Claude Sonnet to decide mouse clicks and key presses. The agent currently runs in a continuous loop. How should I structure the screenshot capture and loop mechanics to prevent high latency and runaway loops? | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-13 We want to implement a confirmation gate class for a desktop AI agent to pause sensitive operations. How should actions be categorized into severity levels, and which levels require mandatory user confirmation? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-01 I want to deploy a Claude-powered screen automation bot to perform GUI tasks on Linux, but I need to isolate it safely so it can't harm my local machine. Could you write a complete Docker configuration and Python launcher script that sets up a virtual display environment with strict container restrictions, resource caps, and network constraints? Please outline the full setup and return the container config files along with the wrapper code. | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-04 I am configuring a Python script using the Anthropic API client to run Claude 3.5 Sonnet with the computer tool for desktop automation. When calling client.beta.messages.create, my requests fail or don't recognize the computer tool. What specific parameter or header flag must be supplied to enable computer use? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-15 I am writing a docker-compose.yml file to run a desktop automation agent container securely. What security options, volume configurations, and resource constraints must be declared in docker-compose.yml to enforce sandbox isolation? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-16 Web security systems flag our GUI automation agent because the timing between actions is completely uniform and exhibits zero mouse movement during thinking phases. How can we humanize action timing in Python? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-18 When our GUI agent attempts to scroll down a scrollable panel, it often fails quietly without moving the view. How should the agent verify whether a scroll action actually succeeded? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
▸case-19 We have a Python wrapper class that starts a Docker container for running GUI automation tasks. If an agent gets stuck in an infinite loop or stalls inside the container, how should the Python wrapper ensure the container terminates automatically after a maximum duration? | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |