A Nvidia study shows that AI agents can maintain accuracy on complex tasks by periodically condensing their history into handoff notes. The researchers reduced the active context to about 16,000 tokens while matching the performance of agents using the full conversation history. On the SWE-bench Verified benchmark, Claude Opus 4.5 scored 70.2% with a handoff note after every action, compared with 70.0% using the full history, and reached 74.8% after tuning. This approach could substantially reduce inference costs, since using 256,000 tokens can cost up to seven times more than using 16,000 tokens. The researchers argue that effective context management, based on summarization, selective information retrieval and structured handoffs, may work better than continually expanding context windows.
Source: Read the original article
