🏆 AA 模型能力榜

Desktop-first agent that executes tasks in a sandboxed Linux VM with access to local files. The agent can edit and create local files, and has pre-set skills for handling common file types and general work tasks (e.g. editing Excel spreadsheets). Connects to external tools via computer use, MCP connectors, or browser extension.

Agentic system within the ChatGPT app, powered by GPT-5.4 Thinking/Pro — OpenAI's frontier model for agentic workflows. It runs on a cloud VM with no direct file access. GPT-5.4 brings native computer-use capabilities and desktop-level automation, designed for longer-horizon tasks with up to ~30+ minutes of autonomous execution. Successor to Operator and combines Codex-level coding and Deep Research into one experience.

Open-source, self-hosted agent platform that turns existing chat apps (e.g. WhatsApp, Slack, Teams) into autonomous AI assistants. Also has macOS menu bar companion app and WebChat interface. Tasks run in the background with results in chat. Supports community Skills marketplace, NVIDIA NemoClaw stack for secure scaling, and security models.

Open-source, self-hosted agent platform with persistent cross-session memory and a self-improving skills system. Creates procedural skills from experience and reuses them across sessions. Tasks run in the background with results in chat. Includes 40+ built-in tools (e.g. image gen, TTS, vision, Telegram, Discord, Slack, WhatsApp).

Agents run inside Microsoft 365 ecosystem (Teams, Word, Excel, PowerPoint, and Outlook), executing multi-step tasks like document creation, data analysis, and scheduling. Connects to external tools via MCP, Power Platform connectors, and Microsoft Graph. Copilot Studio lets organizations build and deploy custom autonomous agents.

No-code agent builder inside Google Workspace for creating and sharing AI agents. Agents reason, adapt to new information, and handle tasks like email summarization, meeting action items, and document management, with deep Google ecosystem integration (Gmail, Drive, Sheets, Maps, Docs, etc.) and external app support via Marketplace connectors.

Agentic platform for enterprise knowledge workers. Connects to company data via prebuilt connectors (M365, Salesforce, Jira, Confluence, Box). Offers permission-aware search, prebuilt agents like Deep Research, no-code workbench, and Code Assist.

General purpose agent that can orchestrate multiple AI models per task and autonomously plan and execute (web automation, coding, data work, PDF/PPT generation) in a sandboxed environment. Produces deliverables including research reports, websites, and data analysis.

Execution layer for Microsoft 365 Copilot that handles autonomous, multi-step actions across Outlook, Teams, Excel, Word, PowerPoint, and Work IQ. Fully sandboxed your M365 cloud tenant, permission-aware, and enterprise-compliant. Plans, coordinates, and executes tasks (calendar cleanup, competitive research, launch plans, etc.) in the background with checkpoints, clarification requests, and user approvals.