How do you control context usage and provider limits when working with AI agents?

I try to treat context as a limited working set, not as a permanent transcript.

A practical setup is:

• keep the current goal, constraints, decisions, and next action in a short task state;
• store large documents and artifacts outside the prompt and retrieve only relevant pieces;
• summarize completed work into durable notes;
• set token and time budgets before starting;
• compact or restart when the context becomes mostly logs, failed attempts, or obsolete branches.

The reason is that a larger context window does not guarantee better use of information. The “Lost in the Middle” study found that models can use information near the beginning and end of a long context more reliably than information buried in the middle. Anthropic’s guide to context engineering similarly recommends curating a minimal useful set of context and tools.

For provider limits, I would track input tokens, output tokens, retries, concurrency, and rate-limit events separately. Then route or queue work according to the actual bottleneck. A compact state plus external memory is usually more reliable than keeping every previous message visible forever.