Taxonomy of autonomous agents
Seven orthogonal axes. Each agent is a vector; comparing two agents is seeing which coordinates differ. Governance risk is a function of the coordinates, not the brand.
To place an autonomous agent you don't need its brand — you need its coordinates. This taxonomy encodes each agent along seven orthogonal axes, so comparing two systems reduces to spotting where their vectors differ, and governance risk becomes a function of those coordinates rather than the marketing. It maps ~18 agents from every major lab plus the open-source frontier, keeps the harness layer deliberately apart, and reads the market's shape: which cells the labs occupy, which one only open source dares, and where the governance line falls.
Filter by axis
| Agent | Vector | Surfaces | Status | Risk |
|---|---|---|---|---|
OpenClaw Open source | A4T3D2M3L2I1 | OS · API/MCP · Chat | Open source | Critical |
Hermes Agent Nous Research | A4T3D2M3L2I1 | OS · API/MCP · Chat · Browser | Open source | Critical |
Gemini Spark Google | A3T3D2M2L1I2 | API/MCP · Browser | Beta (US) | High |
Scout Microsoft | A3T3D2M2L1I2 | API/MCP · OS · Browser | Private preview | High |
Daily Brief Google | A3T3D1M2L1I1 | API/MCP | GA (US) | Critical |
Copilot Studio agents Microsoft | A3T3D1M2L2I2 | API/MCP | GA* | High |
Claude Cowork Anthropic | A2T2D2M2L1I1 | OS · API/MCP | GA | Medium |
ChatGPT Agent OpenAI | A2T2D2M2L1I1 | Browser · API/MCP | GA | Medium |
Manus Manus AI | A3T2D2M2L2I1 | Browser · OS · API/MCP | GA | High |
Devin Cognition | A3T2D1M2L1I1 | Repo | GA | High |
Jules Google | A3T2D1M2L1I1 | Repo | GA | High |
Codex OpenAI | A3T2D1M2L1I1 | Repo | GA | High |
Claude Code Anthropic | A3T2D1M2L1I1 | Repo · API/MCP | GA | High |
Amazon Q Amazon | A2T2D1M2L1I1 | Chat · Repo | GA | Medium |
Nova Act Amazon | A2T2D2M1L1I1 | Browser | Preview | Medium |
Kiro Amazon | A2T2D1M2L1I1 | Repo | GA | Medium |
Sierra Sierra | A3T3D1M2L2I2 | API/MCP · Chat | GA | High |
Lindy Lindy AI | A3T3D2M2L2I1 | API/MCP | GA | Critical |
Harness layer — do not classify as agents
A harness is the scaffolding around the model: context, tools, memory, evaluation. Spark = Gemini 3.5 Flash (model) + Antigravity (harness). The distance between A3·M2 and A4·M3 is harness distance, not model distance.