Some Provocative Takes on AI Agents
The conversation around AI Agents is evolving at a breakneck pace. We're moving from abstract ideas to tangible products, and with that comes a necessary shift in how we think about building, training, and aligning them.
I’ve been wrestling with a few ideas that feel a bit like "hot takes" – unconventional, perhaps, but grounded in the practical realities of making Agents work. I’m sharing them here not as final verdicts, but as a starting point for conversation. I welcome your comments and critiques.
Thought 1: Reasoning Models are Students, Agent Models are Professionals
Let's use an analogy. Training a foundational "reasoning" model is like putting a student through school and university. The goal is to build general capabilities. They learn math, physics, literature, and computer science—acquiring a broad toolkit for understanding the world and solving problems in principle.
Training an "Agent" model, however, is like that student's first day on the job. The focus shifts dramatically from general knowledge to specific, practical skills. A bank teller isn't solving differential equations; they're learning to process transactions with speed and accuracy. A doctor isn't just reciting biological theory; they are diagnosing patients based on specific symptoms and histories. The goal is no longer general intelligence, but effective performance in a defined role.
Thought 2: Deconstructing Alignment: Capability vs. Intent
If we break down the "alignment problem," we can split it into two parts: Capability Alignment (Can the model do the task correctly?) and Intent Alignment (Does the model understand and strive for the desired outcome of the task?).
Following the school vs. job analogy, training reasoning models primarily tackles Capability. We're making the student smarter and more capable in a general sense.
When we train an Agent, the focus flips. We are fine-tuning for Intent Alignment with a smaller, more specialized dose of capability on top. We're not just teaching the bank teller how to use the computer system (capability), but ensuring they understand the goal is to serve the customer efficiently and securely (intent).
Thought 3: Many Agent Tasks Don't Require a PhD
A controversial but important point: many, if not most, immediate applications for Agents don't require profound reasoning abilities. Just as many jobs don't require an advanced degree in mathematics to be performed well, many Agent tasks are more about procedure and reliability than deep, multi-step inference.
This might explain why models not explicitly sold as "reasoning-first" can still achieve top scores on Agent benchmarks. These benchmarks often test the ability to follow a specific workflow or tool-use syntax correctly. It's less about being a genius and more about being a diligent and reliable worker.
Thought 4: For Agents, There's No Magic Bullet—Only Practice
If you ask a master craftsperson, a seasoned pilot, or a top athlete how they got so good, you’ll get the same answer: "Practice." There are no shortcuts. The old saying, "practice makes perfect," is the fundamental rule.
AI Agents are no different. If you want to improve an Agent's ability to perform a specific task, the most effective method is to train it relentlessly on data from that exact task's environment. You move from a "rookie" to an "expert" Agent by showing it thousands of real-world examples of the job done right.
Thought 5: The Myth of Universal Skills
In our careers, we love the idea of "transferable skills." We tell ourselves that mastering one thing will help us master everything else. But in practice, this generalization is surprisingly limited. The specific, hard-won expertise of a great bank teller doesn't really help them become a great investment banker or a financial analyst. The domains are too different.
For Agents, the lesson is even starker. The biggest performance gains will always come from "in-distribution" data. If you want an Agent to be great at processing insurance claims, you need to flood it with data on insurance claims. That specific, domain-focused training will beat a more generalized approach every single time.
Thought 6: Prompting is the Worst Way to Build an Agent
Let's be blunt: as a method for creating a truly competent Agent, prompting is the least effective approach.
Imagine trying to teach someone to drive by sitting them in a classroom and having them read the driver's manual cover to cover. You can give them all the rules, principles, and diagrams you want. But without ever getting behind the wheel, are they truly ready for the road? Of course not. Prompting is the driver's manual; it imparts rules, not experience.
Thought 7: Why We Need Multi-Turn RL: Moving from the Simulator to the Street
This brings us to the "why." If prompting is the manual, what’s the actual driving lesson?
Previous training methods, which often involve single-shot, static examples, are like a driving simulator. It's a safe, controlled environment that's useful for learning the basics. But it lacks the stakes and dynamism of the real world.
Multi-Turn Reinforcement Learning (RL) is the equivalent of getting in the car and actually driving on the road, even if it's just a pre-planned exam route. The Agent must make a decision, see the consequence, and then make the next decision based on that new reality. This loop of action, feedback, and reaction is fundamentally different from a simulator. It's how real learning happens, and it's how we'll bridge the gap from Agents that can follow instructions to Agents that can truly handle the complex, sequential nature of real-world tasks.