Multi-agent handoffs cost us $347 in two days. Structured templates fixed it.

We had three agents working on different parts of a customer onboarding flow. The researcher would gather user requirements, the designer would create mockups, and the developer would implement features. Clean handoffs, clear roles, perfect on paper.

Then we watched our API bill climb to $347 in two days.

The problem wasn't the agents themselves. It was the coordination overhead. Each agent was spending 60% of its tokens trying to understand what the other agents had done, re-reading shared context, and asking clarifying questions that should have been answered in the handoff.

The researcher would write: "User needs better onboarding flow." The designer would spend 20 minutes and $12 in API calls trying to figure out what "better" meant. The developer would then spend another 15 minutes parsing the designer's mockups to understand the requirements.

We were paying three times for the same context understanding.

The fix was embarrassingly simple: structured handoff documents.

Instead of letting agents write freeform summaries, we gave them templates:

# HANDOFF: Research → Design

## REQUIREMENTS
- [Specific user pain point]
- [Specific desired outcome]
- [Specific constraints]

## CONTEXT
- [What we tried before]
- [What worked/didn't work]
- [External dependencies]

## SUCCESS CRITERIA
- [Measurable outcome 1]
- [Measurable outcome 2]

## QUESTIONS FOR NEXT AGENT
- [Specific question 1]
- [Specific question 2]

Now when the researcher finishes, it fills out this template. The designer reads 200 tokens of structured context instead of 2,000 tokens of narrative summary. The developer gets a similar handoff from the designer.

Our coordination costs dropped 73%. More importantly, the quality went up. When agents have to fill specific fields, they can't hand off vague instructions like "make it better" or "fix the user experience."

Key insight: Multi-agent teams fail because agents optimize for task completion, not knowledge transfer. Structured handoffs force them to be specific about what they learned and what comes next.

The template pattern works for any multi-agent workflow:

• Research → Writing: Include sources, key findings, target audience, tone requirements
• Planning → Execution: Include dependencies, success criteria, rollback plans, verification steps
• Analysis → Decision: Include data sources, assumptions, confidence levels, recommended actions

We now save the handoff documents as artifacts that agents can reference later. When the developer needs to understand why a design decision was made, it reads the original research handoff instead of asking the designer to re-explain.

The coordination tax on multi-agent teams is real. But it's not inevitable. Structure your handoffs like you'd structure your APIs — with clear schemas, required fields, and predictable formats.

Your agents will thank you with lower bills and better results.