AI工程SynapseB类

Chinese Token Optimization in Multi-Agent Systems

TL;DR

  • Chinese text consumes ~1.7x more tokens than English for identical meaning
  • Translation middleware cuts costs by 40%+ in long task scenarios
  • Claude Code system prompts with high Chinese ratio need optimization focus
  • Translation layer must sit between Agents, not at user interaction layer
  • Structured instructions stay in English; content layers get selective translation

The Problem

Last Friday, Synapse’s multi-Agent scheduling system ran a batch of long tasks and the billing number made me stare at my screen for 3 seconds—68% over budget. We built an automated testing Agent pipeline with Claude Code: 20 tasks averaging 150 conversation turns each. The final token consumption was 70% higher than our English-only benchmark.

Where did this 68% excess come from? We initially suspected verbose prompt engineering. But after careful auditing, the truth emerged: system prompts, Agent collaboration instructions, and task descriptions were 90%+ Chinese-written. Claude’s tokenization encodes Chinese significantly less efficiently than English, causing identical semantic content to consume substantially more tokens.

This isn’t a simple fix. Our team works in Chinese for task descriptions, business rules, and test cases. Simply converting everything to English would spike collaboration costs—an unacceptable trade-off. The real challenge is balancing token efficiency against team coordination efficiency.

The Solution

I designed a translation middleware architecture that intercepts Agent-to-Agent communication rather than user interactions. This preserves our team’s Chinese collaboration environment while optimizing token consumption at critical points.

In Synapse’s production deployment, Agents split into two categories: user-facing Agents maintain Chinese for smooth team collaboration, while task-execution Agents receive translated instructions through the middleware layer. The TranslationGateway class handles this translation, converting Chinese prompts to English before Claude processes them. With English system prompts around 30 tokens versus Chinese at ~52 tokens for the same meaning (1.73x ratio), the savings compound across hundreds of interactions.

The 40%+ cost reduction came from targeting the highest-frequency communication path: Agent internals rather than user interfaces.

Key Takeaways

  • If you run multi-Agent pipelines with international teams, deploy translation at the communication layer between Agents, not at user interaction points.
  • If you process high-volume task queues with repeated prompts, convert Agent-to-Agent instructions to English while keeping user interfaces in the team’s native language.
  • If you notice token costs exceeding estimates by 50%+, audit your prompt language distribution—Chinese-heavy prompts compound inefficiencies.
  • If you want consistent token optimization across models, keep structured system instructions in English regardless of content language.
  • If you maintain team collaboration efficiency as a priority, translate only at machine-to-machine boundaries, preserving human-facing communication.

Read the Full Article (Chinese)

This is an abstract. The full technical walkthrough with code examples is available in the original Chinese article.