Token Efficiency in Agent Protocols: What to Measure Before Scaling

Most teams measure the final answer and ignore the conversation that produced it. Multi-agent systems make that mistake expensive. Planner messages, tool schemas, role reminders, state summaries, critiques, retries, and review loops can consume more tokens than the final response.

A protocol benchmark should separate task understanding, delegation, tool access, verification, and final synthesis. Otherwise the team cannot see which layer is wasteful.

For Agora, measure how many messages are required before agents agree on responsibility. Then measure how compactly evidence and uncertainty move across the protocol. A good Agora flow should make handoffs clear without carrying every internal thought.

Also test refusal and recovery. Efficient systems do not say yes to bad tasks and spend hundreds of tokens failing. They reject, ask for missing context, or reroute.

For MCP, measure server descriptions, tool schemas, resource listings, and the prompt context needed to use tools safely. Tool access can be standardized and still verbose.

The fix is scope. Agents should see only the tools and resources needed for the current task, not the entire organization.

LangGraph should be measured on state serialization. CrewAI should be measured on role prompt repetition. AutoGen should be measured on turn count and debate loops. LlamaIndex and Haystack should be measured on retrieved context size and citation precision.

Observability tools matter because they turn these guesses into traces. Without traces, token efficiency becomes folklore.

Create a token budget for every benchmark task. Track coordination tokens, tool tokens, retrieved context, review tokens, and final answer tokens separately. Publish the method even when the numbers are early.

This is how comparison guidance can stay unbiased: recommend the stack that performs best for a use case, not the one with the most exciting story.