Optimizing Context for Tool-Using AI Agents

This article discusses the importance of curating context for tool-using AI agents, rather than simply providing a large context window. It outlines four distinct phases of agent execution and the specific context needed for each phase.

Details

The article argues that the real problem for tool-using AI agents is not the capacity to handle large context windows, but the curation of relevant context. Many agents simply concatenate everything - conversation history, tool catalogs, raw outputs - into a single prompt, leading to higher costs, latency, and worse decisions as useful context gets buried in noise. Instead, the author proposes that agents have four distinct phases (route, call, interpret, answer) where each phase requires different context. For example, the routing phase only needs a compact view of available tools, while the answer phase requires relevant turns and the dependency chain, but not the full conversation or raw tool payloads. The author's library, Contextweaver, treats context assembly as a compilation problem, selecting the minimum useful context for each phase within a fixed budget, preserving dependencies, filtering oversized payloads, and deduplicating overlapping context.