Hit 40,000 tokens of context and my AI agent finally stopped losing track

I was testing a multi-step research agent that kept forgetting what it was doing around the 15k token mark. Last week I bumped the context window up to 40k and suddenly it carried a full 12-step workflow without dropping a single instruction. The kicker was the memory usage, it only added about 20% more compute time but the output quality jumped way more than I expected. Has anyone else found a specific context size where their agent just clicks, or is it more about how you structure the prompts?

1 comments

1 Comment

the_viola14d ago

Bumping from 12k to 32k did the exact same thing for mine, night and day difference.