Long-running agents do not just need a bigger context window. They need to learn what deserves to stay in context at all.

Long-horizon agents work better when context management is a learned policy

Long-running agents keep adding searches, tool outputs, and reasoning to the prompt. Eventually the model spends more tokens trying to separate useful facts from old clutter.

ContextPilot gives the agent control over that problem. It can plan, save important information to long-term memory, and summarize, compress, or remove old context.

Its RL training then teaches which of those context decisions actually help.

With Qwen3-8B, ContextPilot-8B-RL averaged 69.40 across 4 long-context benchmarks, versus 45.93 for the same model using a 128K window without context tools.

On BrowseComp, its context stayed around 8K–10K tokens per turn while WebExplorer-8B grew toward 30K.

Paper's overall recommendation: stop treating the entire conversation history as memory. Give agents a smaller working context they can actively manage, and train them to keep what matters.

– arxiv. org/abs/2608.28476

Title: "ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL"