As agents evolve, prompts get unruly. Edge cases are patched with one-line hotfixes, or with stacked instructions of escalating severity when an issue proves stubborn. Over time, those instructions pile up into a brittle artifact, overfit to an aging model. This is prompt debt, and it’s common.

The problem gets worse when agents start managing their own contexts, accruing history through tool calls, user input, and more. Million-token windows aren’t a panacea: they leave more room for bad information as well as good, and quality degrades long before the window fills.

This talk starts with a deliberately simple agent that patches a bug in a repository, then breaks it in four predictable ways: context distraction, confusion, clash, and poisoning. For each one, we add a small, concrete fix: compacting history, selecting only the tools a task needs, and offloading context to a code environment the agent can explore on demand. You’ll see what each fix changes, what it costs, and when it isn’t worth it. You’ll leave with a way to diagnose context failures in your own agents and coding tools, a vocabulary for naming them, and a repo with the code to get started.