The framework wars taught us that the winning layer often does less. The same pattern is emerging for AI agents, where narrow agents need meta harnesses above them.

Drawing a parallel to the JavaScript framework wars of the 2010s, this piece argues that AI agents will follow the same evolutionary pattern: narrow, focused agents will win over batteries-included ones, and a new orchestration layer — the 'meta harness' — will emerge above them. Just as React's deliberate minimalism beat AngularJS, and Next.js filled the gaps React left, today's winning agents will stay focused on their core loop while a meta harness handles cross-repo coordination, persistent memory, multi-agent orchestration, CI integration, guardrails, and more. Most organizations are already building these layers internally in ad-hoc ways, and the author predicts general-purpose meta harnesses will emerge as the dominant pattern.

Meta Harnesses, Agents, and Lessons from the Framework Wars

I’m working on an orchestration layer that supports Claude Code, Codex, Gemini, Hermes, OpenClaw, and Pi, and what you describe in your article resonates a lot with my current experience.
I feel that the ‘I support all tools, all skills, etc.’ approach some harnesses are taking is the wrong way to go and it will end badly. Simplification and specialization are needed, just as human jobs are specialized. An agent that tries to do it all will mess up the context, eat max tokens unnecessarly, etc.
A ‘framework’ that wraps all agents will emerge sooner or later as a standard in the industry, as right now a lot of tweaking is needed to coordinate the different structures gaining traction.
And I also feel we’re not really attacking the real problem. We’re trying to make agents use human tools. But this tools are not the best suited for agents, they were designed for humans. This will evolve into a lot of new tools especifically designed for agents, the web itself blocks lot of bot activity creating bottlenecks everywhere…