Context is a Budget
The more I work with APEX AI Agents, the more convinced I am that context should be treated as a budget. The system prompt, tool definitions, conversation history, and tool responses all spend some of it.
It is tempting to throw everything but the kitchen sink into context and let the LLM figure it out. I think deciding what belongs there is part of our job as APEX developers.
Focused, relevant, high-quality context can improve instruction adherence, factual accuracy, reasoning quality, and performance.
For APEX agents, that means building tools with useful filters that return only the data needed. If someone asks which tasks are overdue, return the matching tasks with their IDs, titles, owners, and due dates. If there are too many matches, return a page of results and make it clear that more are available. Give the agent a tool to retrieve more detail when needed.
Start with SQL filters. Use Oracle Text, ubiquitous search, vector search, or fuzzy matching when the question calls for them.
Give the agent what it needs now, with a way to find more when it needs it.
A bigger context window gives us more room. It does not remove the need to spend it carefully.
