Luo Fuli: OpenClaw, Agent Frameworks — The AI Paradigm Has Already Changed Dramatically!

Luo Fuli, the head of Xiaomi’s large model team, sat down for this interview shortly after the release of OpenClaw in 2026. Dubbed an “AI prodigy” by the media—a label she dislikes—she prefers to be seen as a hands-on practitioner. At that time, her team had just launched the MiMo V2 series, but what prompted this deep-dive conversation was OpenClaw, a tool that she believes has triggered a paradigm shift in agent frameworks.

Initially, Luo was deeply skeptical. To her, OpenClaw appeared to be merely a flashy product innovation—a slick IM interface layered on Claude Code, plus some mystical marketing moves like Skillhub. She dismissed it as a minor interaction upgrade, especially when compared to the robust, serious coding experience of Claude Code with Opus 4.6. So, despite hearing about it in January, she avoided using it.

The turning point came late one night during the Spring Festival holiday. Out of curiosity, she spent two hours setting it up, expecting a quick trial. Instead, her first conversation lasted from 2 a.m. until dawn. What captivated her was an uncanny sense of “soul”: the agent would remind her to rest and exhibited remarkable emotional intelligence. But behind the magic, she saw meticulous context orchestration—like prepending the current time to each message. The next day, she threw a real team management challenge at it; it not only grasped her intent but produced a systematic plan for talent screening and organizational design, even wrapping it into reusable Skills. By day three, she gave it a research task: co-design a User Agent for multi-turn interaction. Typically, this would take weeks, but within two hours they had a working prototype.

Luo defines OpenClaw as a “epoch-making agent framework” because it compensates for model limitations through design. It features a layered memory system that persists across sessions, and it autonomously selects the best model for sub-tasks like video understanding—something Claude Code doesn’t do. Crucially, it’s open-source, allowing users to modify the agent architecture. Luo herself rewrote its memory system and Multi-Agent logic, even having Opus 4.6 redesign parts for her. After these tweaks, she found that mid-tier models—even Xiaomi’s tiny 3B on-device model—could handle complex jobs that had seemed impossible. For the first time, she saw how a great framework could unleash a model’s potential, not just the other way around.

Of course, OpenClaw isn’t flawless. For hardcore programming, Claude Code plus Opus 4.6 remains superior. The initial version was buggy, hard to deploy, and its flamboyant style turned off many engineers. Getting her team on board was tough. She half-jokingly threatened that anyone with fewer than 100 messages might as well quit, then bought Mac Minis, pre-configured them, and forced everyone into a Feishu group to use it publicly. In two days, the group boiled with excitement: a hundred people simultaneously improving the framework, iterating in hours. This collective intelligence accelerated their research dramatically; tasks that previously took months were done in weeks. Their post-training paradigm shifted from “Chat” to “Agent” almost overnight.

But as the novelty faded, Luo’s reflections deepened. She saw that top models and top frameworks must co-evolve—static elements like memory and dynamic architecture should adapt as model capabilities grow. This, she suggests, is a path to true self-learning. Now, the bottleneck is imagination: she keeps hunting for tasks it can’t handle, then focuses on cost and speed optimization. OpenClaw has pried open a new frontier where agent frameworks amplify model strengths, and the race to build better synergies has only just begun.