The Dependency Graph of Tools: A Scalable Architecture for Agents

In our previous lessons, we built an agent capable of selecting the right tool for a given job. But we made a crucial simplifying assumption: that all our tools are independent. As we scale our systems to solve real-world problems, this assumption breaks down, revealing a critical vulnerability in standard agentic design.

The standard Retrieval-Augmented Generation (RAG), whether used for retrieving documents or for selecting tools, operates in a “flat world.” It treats every item in its knowledge base (be it a document chunk or a tool description) as an independent entity stored in a vector space.

This approach is powerful for finding items that are semantically similar to a query. However, it is completely unaware of the explicit, structured relationships between these items. This unawareness leads to two major failure modes in complex, real-world scenarios.

Limitation 1: Ignoring tool dependencies

In any sophisticated system, tools are not standalone; they have dependencies. One tool often requires the output of another to function correctly.

• A get stock price tool requires a ...