SettledResearch
Microsoft study: 7.8% of Copilot coding sessions need compaction, using 44% of tokens
Microsoft researchers analyzed 13.5M GitHub Copilot agentic coding sessions (760M LLM calls) and found 7.8% triggered context compaction, yet those sessions used 44.2% of all tokens. Their method, APEC, prepares summaries in the background and cuts cost by up to 25.4% and run time by 9.6% across frontier models, they say, with better task quality. They also release CompactionBench.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Microsoftmicrosoft.com/en-us/research/publication/benchmarking-long-horizon-agent-context-compaction-with-compactionbench/Coverage · 1 article, oldest first
Friday, November 13, 2026