Stop managing agents by context-window length. Manage them by token count.

Here's the cheapest performance fix in the building, and most teams skip it because it sounds boring. Databricks observed correctness degrading around 32k tokens even on large-window models. So a 200k window is not 200k of usable reasoning. It's maybe 32k of good reasoning and a long tail of "context rot" after that. Separately, Anthropic found token usage explains roughly 80% of multi-agent performance variance. More than model choice. More than tool-call count. (Digital Applied)

Read that again. The single biggest predictor of whether your agent succeeds is how many tokens are in play, not which model you picked. That inverts how most people debug. When an agent fails, the reflex is "let me try a bigger model" or "let me upgrade to the new frontier release." The data says the higher-leverage move is to compact your context first.

HumanLayer's ACE-FCA playbook puts a hard number on it: keep coding-agent context utilization between 40 and 60%, and force intentional compaction before you cross ~60%, around 170k of a usable window. Above that, output quality degrades. The compaction step distills progress into a structured doc (goal, approach, completed steps, blockers) and restarts on a fresh window. (HumanLayer ACE-FCA) One practitioner shipped 35k lines into a 300k-line Rust codebase in 7 hours using exactly this checkpoint discipline.

This connects straight to the Anthropic 400K study. Experts write 3,200-word prompts and trigger 12 actions because they're managing context deliberately, not because they type fast. The skill isn't prompting. It's deciding what the model should and shouldn't be holding in its head at each step.

The action is concrete and you can do it today. Add a tokens-in-play counter to your agent loop. Treat it as your primary health metric, the way you'd treat error rate or latency. When you cross ~60% utilization, compact to a structured progress doc and restart on a clean window before you reach for anything bigger. I've watched agents go from confidently wrong to correct on the same task just by clearing the cruft out of the window. The model was never the problem. The clouded context was.