Context Optimization

When you're hitting context limits or watching token costs spiral, this gives you four concrete levers: compress old conversation turns into summaries, mask verbose tool outputs with compact references, optimize for KV-cache hits by reordering stable content first, or partition work across sub-agents with isolated contexts. The claim is you can double or triple effective capacity without quality loss. Most valuable when tool outputs dominate your token budget or you're building long-running agents. The compaction and masking strategies are immediately practical. The real insight here is that context quality beats quantity, so you're strategically choosing what to keep rather than trying to cram everything in.