Six patterns cut agent token use 60-70%

A production dashboard cut agent cold-load times from a 3.5-10.5 minute baseline to 61-116 seconds using six token-saving patterns, per arXiv 2608.17188. The paper estimates the combined set cuts token use by 60 to 70 percent.

That number lands hardest on anyone writing an MCP server. The paper's fetch-once/process-locally pattern names a specific failure: a paginated endpoint that forces the calling agent into six round trips. Each round trip bills tokens to whoever installed the tool, not its author.

The other five patterns: context stratification, schema-contracted prompts, token-aware fallback chains, semantic caching, and inter-agent communication compression. None are exotic. They read like fixes for problems anyone watching a token bill would eventually notice.

Semantic caching and schema-contracted prompts solve a related problem. Don't make the model re-derive structure it already produced. Don't let it guess at a response shape when one can be pinned down.

Token-aware fallback chains are different: they're an admission that agents run out of budget mid-task, and a graceful downgrade beats a hard failure.

The paper doesn't say how these patterns hold up outside the dashboard it was tested on. It also doesn't say whether the 60 to 70 percent token savings generalize to different fetch patterns. The practical check is cheap regardless: count how many round trips your MCP tool forces per query, and ask who's paying for them.