Context Engineering
for Developers

Context breaks before it overflows. Long-running agent sessions degrade in predictable ways, and most developers don't notice until the output has already gone to garbage.

So context rots. The next question is whether you can delay that by being smarter about what goes in and where you put it.

Good structure and careful selection buy you time, but context still grows with every turn. At some point the window fills up, and you need a plan for what to compress, what to persist outside the window, and what to let go.

You can structure context well and manage its growth, but you're still paying full price for every file the model reads, every single turn. One pattern changes that math.

You've got the right context in the window and it's not growing out of control. Now the question is whether the model actually uses what you gave it, or hallucinates around it.

The core patterns apply regardless of what you're building, but the specifics of applying them differ by domain. Pick whichever guide is closest to your use case.

This agent has six context engineering problems stacked on top of each other. Find them all, name the pattern that fixes each one, and rewrite the configuration from scratch.