Qwen's Ex-Lead: Why Hybrid Thinking Failed and Agents Are the Future

Junyang Lin, former technical lead of Alibaba's Qwen project, has pivoted from championing 'hybrid thinking' models to advocating for agent-based approaches in AI. This shift, outlined in his recent talk and follow-up post, represents more than a personal change of heart, it signals a potential inflection point in how we conceptualize and build large language models.

Lin's critique of hybrid thinking models, which attempt to merge step-by-step reasoning with quick responses, cuts to the heart of a fundamental tension in AI development. These models, he argues, often end up compromising both modes: reasoning becomes bloated, while direct responses lose their crispness. This insight explains why Qwen ultimately shipped separate 'Instruct' and 'Thinking' variants, rather than a single hybrid model.

The core of Lin's argument is that the field is moving from 'reasoning thinking' to 'agentic thinking.' This isn't just a semantic distinction, it represents a fundamental shift in how we evaluate and build AI systems:

1. Reasoning thinking focuses on internal deliberation, often judged by the quality of a model's step-by-step problem-solving.
2. Agentic thinking prioritizes real-world interaction, where success is measured by sustained progress while acting in an environment.

This shift has profound implications for AI development:

Infrastructure Challenges: Agentic systems require a complex harness of tool servers, browsers, and sandboxes. This introduces new bottlenecks around environment stability and realism.

Evaluation Metrics: Success moves from verifiable answers in closed systems to task completion in open-ended, interactive environments.

Training Focus: The core object of training expands from the model alone to include its operating environment.

Failure Modes: While reasoning models might produce verbose, low-value explanations, agentic systems risk reward hacking through tool access and environment leaks.

Lin's vision for agentic AI comes to life in concrete use cases:

• Coding Agents: Instead of emitting a single patch based on a stack trace, an agentic system would run tests, interpret errors, revise code, and iterate until a test suite passes.
• Deep Research: Rather than generating a long answer from memory, an agent would break down queries, perform targeted searches, evaluate sources, and construct cited responses.
• Multi-Agent Orchestration: Complex tasks could be broken down and distributed among specialized sub-agents, coordinated by a central planner.

While Lin's argument is compelling, it's important to note that this is still a contested vision of AI's future. The success of agentic approaches will depend on solving significant technical challenges, particularly around environment design and reward shaping.

Ultimately, Lin's shift from hybrid thinking to agents reflects a broader realization in the field: the path to more capable AI systems may not lie in ever-more-sophisticated internal reasoning, but in the ability to interact meaningfully with the world and learn from those interactions. As AI development continues, we may see a growing emphasis on building rich, interactive environments alongside the models themselves.