When an agent performs badly, one of the most common fixes is to give it more context.

Add more documentation. Retrieve more files. Increase the context window. Include more history. Attach the entire repository. Keep every previous tool result.

Eventually, the agent knows everything but understands nothing.

Large context windows have created the impression that context scarcity has disappeared. In practice, a different constraint remains: attention.

This session proposes a different architecture for agent context.

Instead of treating context as a bucket, treat it as a dynamically assembled working set: selected according to the current task, decision and stage of execution.

We will examine contextual scopes, progressive disclosure, retrieval boundaries, ephemeral tool outputs, repository navigation and strategies for letting agents fetch information when needed rather than carrying everything everywhere.

This session gives engineers a more useful way to think about the problem: not “How do I fit everything?” but “What does the model need to decide next?”

Target audience: Engineers building coding agents, RAG systems, autonomous workflows, developer tools or AI-assisted products.

Preferred duration: 35–45 minutes.

Technical level: Intermediate.

Format: Architecture patterns, failure examples and practical context-selection strategies.

Framework-independent: Applicable across coding agents, browser agents, RAG applications and tool-using systems.

Prerequisites: Basic familiarity with LLM applications is sufficient.