Agent Tool Use

Tool use is how an agent acts on the world, and the literature organises it as an evolutionary progression through three paradigms — prompting, supervised training, and reinforcement learning. The progression is one of where the competence lives: instructed at inference, trained into the weights, or learned from outcome feedback.

Agentic reasoning more broadly resolves into a three-layer frame: foundational, self-evolving, and multi-agent. The layers are frequently conflated in product claims — a system described as "multi-agent" is often a single agent with foundational tool use — and the distinction matters because the failure modes differ at each layer.

The strongest existence proof of the payoff is AlphaEvolve, a Gemini-powered evolutionary coding agent that recovered 0.7% of Google's compute and produced novel mathematical results. It is worth being precise about what that demonstrates: an agent operating a tool loop against a verifiable objective, over a long horizon, at production scale. Where the objective is checkable, the loop is the product.

Key Claims

• Three tool-use paradigms — prompting, supervised, RL — as a unified evolutionary framework. Evidence: moderate (paper) (Agentic Tool Use in LLMs)
• Three-layer agentic reasoning frame: foundational, self-evolving, multi-agent. Evidence: moderate (paper) (Agentic Reasoning for LLMs)
• AlphaEvolve recovered 0.7% of Google's compute via an evolutionary coding agent. Evidence: moderate (technical-report — vendor) (AlphaEvolve)