what makes a good agentic model?

i will start with a counterintuitive hypothesis.

what if models already know how to use tools, and we are just overthinking the training process?

let’s think about it a bit. during pre training, these models consume massive of code. billions of lines spanning across different languages, frameworks and paradigms. every api call in that code is essentially a tool use pattern. when you use requests.get() in python, that’s the model learning how you interact with external systems.

and we know models can role play convincingly. so it might be the case that we don’t need to teach models tool use from scratch but focus more on unlocking the latent ability?

we have seen the from the likes of chinese open sources models. especially kimi one.

instead of laboriously collecting thousands of human annotated examples of tool usage, let the model generate its own training data.

using a multi agent setup, you can have models synthesize diverse tool call scenarios - different apis, different use cases, different interaction patterns. the model essentially bootstraps itself, creating the exact kind of training data needed to activate its dormant tool-use capabilities.

and apparently? it works surprisingly well.

a good agentic model can:

• maintain conversation context

• decide when it needs a tool vs when it can answer directly

• make tool calls with proper params

• incorporate tool results back into its reasoning and know when its done

though i guess we are still early. it’s not about just calling tools but more about handling complex trajectories that weave together exploration, reasoning, environmental feedback, error recovery and adaptive planning. maybe its a model that can:

• explore when its uncertain

• backtrack when it hits dead end

• learn from failed attempts

• adjust its strategy based on what it discovers

i guess the training data might be tricky. synthesising this kind of data is harder, but not impossible. as long as we can decompose complex tasks into these interleaved patterns, i guess we can stimulate the capability in models.

maybe the real capability hiding in these models isn’t tool use at all. maybe it’s something more fundamental - the ability to maintain uncertainty, to explore without knowing where you’re going, to build understanding through interaction rather than execution.

i guess the tools are just the interface. there is more agency in how the model thinks about using them.