Most of the harnesses on this blog live in your terminal or your editor. OpenClaw takes a different starting point: what if the harness lived everywhere you already message people, so the agent could work through the channels you use anyway rather than a dedicated app?
What’s actually under the hood
OpenClaw embeds createAgentSession() from @earendil-works/pi-agent-core – the SDK behind Pi, a minimal terminal-coding harness – directly into its own process, rather than shelling out to a separate Pi CLI. On top of that runtime it bolts on more than a dozen messaging integrations: WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, Matrix, Feishu, LINE, WeChat, QQ, and voice, all from one running service. The trade-off for that breadth is opinionation – OpenClaw makes a lot of decisions about how the agent behaves so you don’t have to wire each channel up yourself.
Agent runtimes and coding harnesses, side by side
For coding specifically, OpenClaw supports two distinct paths. Embedded harnesses – the built-in openclaw runtime plus registered plugin harnesses like codex and copilot – run inside OpenClaw’s own prepared agent loop. External harnesses connect over the Agent Client Protocol (ACP): Claude Code, Cursor, Copilot, Droid, OpenCode, Gemini CLI, and other ACP-compatible tools can be spawned as background tasks and tracked from the same interface. That split matters in practice – selecting an openai/* model, for instance, routes you through the Codex app-server runtime rather than a generic API call, and OpenClaw’s own documentation is explicit that ACP is the external-harness path, not the default one.
Security is the recurring conversation
OpenClaw’s rapid growth – hundreds of thousands of GitHub stars – came with a real reputation for security flaws, enough that industry coverage has used it as the example of what happens when a harness gets popular faster than its attack surface gets audited. A permanently running service bridging a dozen messaging platforms is, by definition, a large surface. If you’re evaluating it, that’s the first thing to read up on rather than the feature list.
Why it matters beyond itself
OpenClaw’s rise is a big part of why “harness” became the industry’s word of choice in 2026 for the code wrapped around an LLM – the orchestration, tool-calling, and context management layer that turns a model into something that gets things done. Anthropic’s own Claude Code plugins remain compatible with OpenClaw-oriented markdown workflows, and several of the projects covered elsewhere on this blog (Munder Difflin, Hermes Agent) explicitly position themselves against or alongside it.
Leave a Reply
You must be logged in to post a comment.