The authors built ctxagent, a local agent harness in which the model manages its own context using two tools: a compaction tool that condenses most of the context at once and a curation tool for smaller deletions.
Letting agents manage their own context can cut cost without hurting performance
Across 35 to 212 turn mixed-task sessions, Claude Opus 5.5 with two context tools matched Claude Code on task success at 34 to 49% lower token cost.
Research Lab
Matthias Busch · Marius Tacke · Roland C. Aydin · Christian Feiler · Christian J. Cyron
Hamburg University of Technology · Helmholtz-Zentrum Hereon · German Research Center for Artificial Intelligence (DFKI) · Saarland University
Research Digest··3 min read
The authors test whether frontier LLMs can decide for themselves when and how to compact their context, instead of relying on fixed token thresholds set by agent harnesses.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.