Korva is a scroll of the internet’s best ideas.
Models get dumber at halfway
How the window fills
A model can only take in so many tokens at once. A common limit today might be 270,000 tokens, and some models have a million. An agent's whole trajectory is made of its own outputs, the inputs from the human, and the inputs from the environment, and every turn of the loop stacks more of them onto the pile.
The dumb zone
After enough turns, the window can literally fill up. Trouble starts long before that. Somewhere around the 50% mark, models get a lot dumber, a stretch people call the dumb zone. The window has not run out, but that is not a place an agent should end up.
How to engineer around the dumb zone is still an emerging practice, and the best teams are doing some interesting things to deal with it.
Why do models weaken in long contexts?
A 2023 study led by Stanford researchers, titled Lost in the Middle, found that models use details at the start and end of a long input better than details buried in the middle. Attention has to spread across every token in the window, so relevant details compete with more and more surrounding text. In 2025, the vector database company Chroma reported accuracy drops as inputs grew across many models it tested.
Why does every loop turn grow the context?
Models keep no state between calls, so each request has to carry the full conversation so far. Every new action, tool result, and reply is appended to that record. The next call then re-reads all of it before adding anything new. Providers offer prompt caching to reduce the cost of re-reading history that has not changed.
How do agents shrink a crowded context?
One common technique is compaction, where the model writes a summary of earlier turns and the harness replaces the long history with it. Another is clearing old tool outputs, such as large file dumps, once they have been used. A third is handing a subtask to a separate agent with a fresh window, which returns only a short result. Claude Code includes a compact command that performs the summarizing step.
More from Lightspeed Venture Partners
