Skills
Compact agent context on a semantic boundary, not a token threshold
A June 2026 'Self-Compacting Language Model Agents' line of work pairs a compaction tool the model can call with a rubric for when to fire (sub-task resolved, trajectory converging) versus hold (mid-derivation, stuck) — making forgetting a semantic decision rather than a 'buffer is full' numeric trigger. Across six benchmarks and seven models it beat a no-summarization baseline by up to 18.1 points on math and 5–9 on agentic search at 30–70% lower cost per question. Builders running long Claude Code / agent sessions should expose compaction as a model-callable action gated by task state, not auto-fire at a percentage.
Source
↳ Follow the thread