Open Agent exposes only 2 tools with minimal schema documentation. Both tools have descriptions, but schemas lack parameter type definitions and descriptions. computer_use_task has a single 'task' parameter with a string type but no constraints or format guidance. computer_screenshot has an empty properties object with no documented output schema. Neither tool provides error handling guidance, recovery paths, or output field documentation. The tool definitions appear inferred from code rather than explicitly registered with full MCP metadata. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present despite computer_use_task being marked IRREVERSIBLE.

• Add explicit output schema documentation for computer_screenshot: e.g., {"type": "object", "properties": {"image": {"type": "string", "description": "Base64-encoded PNG image of the current screen"}, "width": {"type": "integer"}, "height": {"type": "integer"}}}
• Enhance computer_use_task description with scope and format guidance: 'Delegate a task to a computer-use agent. Tasks should be specific, single-goal actions (e.g., "Open Firefox and navigate to example.com", "Click the Save button"). Avoid multi-step sequences; break complex tasks into separate calls.'
• Add destructiveHint and readOnlyHint annotations to tool definitions to enable LLM safety reasoning and prevent accidental misuse.
• Implement error categorization: return structured errors with 'category' (retryable|user_fixable|fatal), 'message', and 'suggestion'. E.g., {"error": "Application not found", "category": "user_fixable", "suggestion": "Ensure the application is installed and accessible."}
• Add optional 'confirm' parameter to computer_use_task (boolean, default false) or implement a separate confirm_computer_task tool to gate irreversible operations behind user approval.