If you’ve ever watched an AI agent demolish a leaderboard only to faceplant on a basic Slack integration, you understand the quiet crisis in agent development. MiniMax’s M2 team, in a candid technical breakdown, argues the field has been optimizing for the wrong thing entirely. The core problem isn’t tool mastery—it’s perturbation tolerance.
The team initially followed the standard playbook: scale up tool variety to teach generalization. Scores climbed. But the moment they swapped the scaffolding framework underneath the model, performance cratered. “We realized we were solving the wrong problem,” the team writes. The breakthrough came from recognizing that an agent’s operating environment is a minefield of constant change—system prompts shift, tool responses are unpredictable, environments mutate. Their old approach only addressed one variable: the tools themselves.
This led to what they call “Interleaved Thinking,” a design where the model’s internal reasoning isn’t a one-and-done preamble but a continuous thread woven throughout the entire task. It lets the agent re-evaluate when a search returns garbage or a code execution throws an error. Professor Junxian He and Wenhu Chen were instrumental in shaping this architecture. The team also issued a practical warning to developers: M2’s context is its memory. Trimming the history—a common habit with simpler reasoning models—silently lobotomizes it, which explains a lot of the performance complaints cropping up in community forums.
What’s striking is the honesty about the gap between lab and life. The post doesn’t just celebrate benchmark wins; it admits that a model acing BrowseComp’s contrived puzzles doesn’t mean it can handle a messy real-world task. The internal results, they claim, exceeded expectations when they threw obscure “cold-start” scaffolding at M2—frameworks they hadn’t explicitly trained for. That’s the acid test most agent papers avoid.
