The Anatomy of the Agent Context Window
Techstrong.ai, Friday, October 16th, 2026
A six-layer model of the agent context window, with techniques to cut token waste and block prompt injection.
Nandish Nanjappa treats an agent's context window as its volatile memory and walks through a corporate flight-booking agent to define six layers.
They are system directives and policy anchored at the start, a skill registry of tool schemas, retrieved ground truth, rolling episodic history, working reasoning and execution state.
Optimizations include dynamically loading only the tools relevant to the current sub-goal, prompt caching and distilling tool observations to save tokens.
For security, he recommends wrapping retrieved content in clear structural tags and strictly separating trusted instructions from untrusted data to prevent prompt injection, data exposure and unauthorized tool calls.