Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Workrun investigates why a simple greeting used 5,671 tokens

AI summary

A Workrun developer reported that the greeting “Hello, how is your day today?” returned normally but consumed 5,671 tokens, prompting an investigation into hidden agent costs. The account describes adding Langfuse observability alongside OpenTelemetry, splitting routing from task execution, using LiteLLM Proxy for team access and usage attribution, and adapting JSON output for a DeepSeek model path. The developer also proposes evaluating workflow requests by routing correctness, usable outputs, verification status, usage, latency, and repair counts rather than token count alone.

In-depth insight

The first-hand Workrun case shows how tracing, routing boundaries, model gateways, and task-specific evaluation can expose and govern hidden costs in AI agent systems.

我原来觉得,Workrun 的对话入口应该有一个足够能干的主 Agent。

用户说什么,它先理解。普通问题自己回答,需要工具就调用工具,需要 Workflow 就交给规划 Agent,方案确认后再交给编排 Agent。复杂性留在内部,用户只面对一个输入框。

直到我发了一句:

你好,今天过得怎么样?

回复没有异常,Workrun 自带的用量统计却显示:5671 Token。

这个数字让我不太舒服。倒不是一次问候的费用有多高,而是我解释不了它。用户没有提出复杂需求,系统为什么要做这么多准备?如果连这笔账都说不清,后面继续加工具、加 Skill、加 Agent,我又凭什么判断成本是合理的?

#Workrun#Langfuse#OpenTelemetry#LiteLLM Proxy#ADK#DeepSeek

Original · 15 min readRead the original
Group chat

No comments yet. Start the conversation.