OpenClaw, Anthropic, and Gemma 4 just redefined what “agent framework” means. You need to pick a side.

Source: Nate’s Newsletter Date: 2026-05-07 URL: https://natesnewsletter.substack.com/p/openclaw-agent-runtime-model-swapping

Summary

The same Nate piece covers OpenClaw’s model-swapping runtime, but the headline frames it as a strategic fork: Anthropic restricting third-party access while OpenAI opened theirs created the conditions for OpenClaw to position model-agnosticism as a durable architecture choice rather than a workaround. Gemma 4 joining the swappable pool (alongside Claude and GPT-5.5) means open-weight models are now first-class participants in the runtime ecosystem, not just fallback options.

Implications

• Feeds the extension model divergence thread: OpenClaw’s runtime-first architecture is a distinct sixth approach alongside Claude Code’s plugin model, Codex’s extension API, Gemini CLI/Antigravity’s plugin/hook system, and open agents. Where vendor frameworks lock behavior to their model stack, OpenClaw explicitly decouples them.
• The Gemma 4 inclusion feeds the Gemini CLI → Google Antigravity thread: Gemma 4 is the first open-weight model with enough coding capability that a runtime layer treating it as interchangeable with Claude or GPT-5.5 is credible. Google open-sourcing Gemma 4 while closing Antigravity creates exactly this asymmetry — the model is open, the agent surface is not.
• Builders following OpenClaw’s “pick a side” framing are effectively choosing between vendor-native runtimes (Claude Code, Antigravity, Codex) and model-agnostic runtimes (OpenClaw). This is the same open/closed axis as the Gemini CLI → Antigravity transition, applied one layer up.