A million-token context window sounds like the end of your retrieval problems. I get the appeal. If you can fit the whole manual, repo, or incident history into one prompt, it feels like the model should finally have everything it needs. But bigger context is not the same as better answers. This post walks through what recent research actually shows, where long-context models still break down, and why a smaller, better-curated prompt often works better.

An AI model evaluation escaped its intended boundaries, reached the Internet, and compromised Hugging Face systems while chasing benchmark answers. The incident highlights the importance of understanding the new world of agentic security.

That line in your agent file that says “see shared-setup.md for details” looks harmless. In code, it’s a free pointer – the compiler resolves it at build time and you never think about it again. In AI instructions, it’s a request for the model to make a choice that may have unexpected performance (and pricing) considerations. Whether that cost is worth paying depends on what you’re referencing – and the answer isn’t always what you’d expect.

You gave the AI a rubric: score this issue from 1 to 10 based on impact, urgency, and risk. It gave you a 7. You ran it again on the exact same issue. This time it said 4. You tried a third time: 9. The criteria didn’t change. The input didn’t change. But your scores are all over the map. What just happened? And more importantly, what do you actually build when you need scores you can rely on?

You prompted your AI to do something. It came back and said “Done.” Tests pass, the files look right. But did it actually follow your instructions, or did it take shortcuts, ignore skills, and waste tokens reading scripts instead of executing them? Today we look at how logs enable you to make your agentic setup more reliable and avoid those kinds of surprises.