OpenClaw

Overview

OpenClaw is a persistent agent runtime optimized for breadth: many integrations, many models, many surfaces, and a large public skill ecosystem. The current docs clarify that this breadth sits on top of a concrete runtime discipline: one main workspace, injected bootstrap files, layered skills, and JSONL session storage. In this wiki, it remains the ecosystem-first counterpoint to hermes-agent.

Integration model

The current docs present OpenClaw as a single embedded agent runtime wrapped by OpenClaw-owned session, delivery, and discovery layers. The required workspace files (AGENTS.md, SOUL.md, TOOLS.md, BOOTSTRAP.md, and friends) make it architecturally interesting even apart from coding: it treats presence, persona, and workspace bootstrapping as first-class harness capabilities. See agent-harness-anatomy, memory-persistence, and instruction-layering.

OpenClaw RL lead

The 2026 self-distillation interview linked by 0xSero mentions OpenClaw RL as a library using on-policy self-distillation under the hood for agents that learn asynchronously from tool and environment feedback. This is a useful lead, not yet a hardened architecture fact for the main openclaw runtime page: the existing primary OpenClaw docs in this wiki cover workspace, skills, and session persistence, not the RL adaptation layer.

Architecturally, the lead matters because it would connect OpenClaw’s ecosystem surface to on-policy-self-distillation, self-evolving-workflows, and evaluation-and-review-loops. It should be verified against primary OpenClaw RL docs or code before being promoted into the formal: block.

Strengths

• Exceptional surface-area breadth.
• Large public skill marketplace and community energy.
• Explicit workspace and session model under the ecosystem story.
• Strong fit for users who want one runtime to be reachable from everywhere.

Risks

The same source also emphasizes serious security concerns: malicious skill entries, supply-chain risk, and unsafe websocket behavior. OpenClaw is therefore a useful object lesson in safety-and-permissions.

Relationships

OpenClaw is best read against hermes-agent, instruction-layering, memory-persistence, and safety-and-permissions. It appears in both harness-quality-comparison and harness-architecture-comparison as the ecosystem-maximal pole.