Multi-agent workflows get a playbook for cutting token waste

A new arXiv paper treats token management as an engineering problem for multi-agent AI systems, not just a billing concern.

The authors describe a production dashboard that uses language models to extract structured work items from meetings, email, and chat, then routes summaries across workstreams. From that system, they outline six patterns: context stratification, fetch-once local processing, schema-contracted prompts, token-aware fallback chains, semantic caching, and compressed inter-agent communication.

In measured runs, the approach reduced cold-load latency from an operational baseline of roughly 3.5 to 10.5 minutes to 61 to 116 seconds, with estimated token reductions of 60 to 70 percent. The paper also reports a controlled context-composition study across 2,420 trials and 11 model configurations.

The important caveat is that smaller context is not automatically better. The work argues for selecting and shaping context deliberately so agents stay cheaper and faster without losing the evidence they need.