· 5 min read
agentic-ai
Stop Letting a Harness Manage Your Agent's Context — Let the Model Edit It Instead
A new paper from the University of Washington, MIT and Meta Superintelligence Labs gives models direct write access to their own context file instead of a hand-built compaction harness. The result on BrowseComp-Plus: 11.4% higher accuracy with 21.5% fewer FLOPs. Here is how the mechanism works, what the numbers actually mean, and the security problem the authors admit they haven't solved.
Read more
