Your agent has the README, the architecture docs, half the issue tracker, and a super prompt that ends with "make no mistakes." A hundred thousand tokens of excellent advice. It still hands back a 500 where the tests wanted a 404.

A bigger context window wouldn't have saved it, because the problem isn't the model, it's the context architecture the model sees.

Baruch will show you four things that quietly wreck that architecture: the stuffed prompt, the wrong tool, the goldfish, and the eval nobody runs. Every one of them is easy to cause and immune to pasting in another document.

Your context window is not a giant socks drawer. Break it into artifacts you can version and test, match the tool to the shape of the problem, and measure what each piece buys you.

And yeah, we'll do it live! It'll be cool if it works and absolutely hilarious if it doesn't.

The session reruns failing agent tasks after each context fix and evaluates the combined changes at the end. Developers can use that approach to test their own context choices. Recorded fallbacks support the live format.