Checked August 2026
Real agent runtime atlas
OpenClaw and Hermes Agent, architecture before hype
OpenClaw and Hermes Agent are open-source agent runtimes, not foundation models. Both can connect models to tools, skills, plugins, persistent state, and messaging surfaces, but they organize the system around different centers of gravity.
OpenClaw is Gateway-first: one long-lived control plane connects channels, clients, paired nodes, sessions, and embedded agent runs. Hermes is AIAgent and learning-loop-first: many entry points converge on one orchestration core with prompt tiers, toolsets, multiple execution backends, bounded memory, and on-demand skills.
- OpenClaw
- Gateway · channels · sessions · nodes · embedded loop
- Hermes Agent
- AIAgent · toolsets · backends · memory · learning loop
- Verified
- Official project docs · August 10, 2026
Gateway-first personal agent
OpenClaw
One long-lived Gateway is the control plane for channels, clients, nodes, sessions, events, and the per-session embedded agent runtime.The documented security model is one trusted operator boundary per Gateway. Mixed-trust tenants should use separate gateways and ideally separate OS users or hosts.
Primary research
Inspect the mechanism at its source.
Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.