
在过去的几个月中,我们都被各种眼花缭乱的新编程 Agent 框架包围了。从 Claude Code 到 OpenCode,再到 mini-SWE-agent,每个开发者生态都想让你相信,他们自定义的框架(Harness)就是能将基础大语言模型(LLM)变成资深工程师的“独家秘方”。但是,如果你在让其中某个 Agent 循环运行修复 Bug 之后查看过 API 账单仪表盘,心里大概会猛地一沉。
最近一项对照研究证实了许多实操开发者在 Discord 频道里窃窃私语的事实:那些花哨的 Agent 框架大多只是在虚高你的云端账单,对实际提升代码质量毫无帮助。
该研究对几种热门的编程 Agent 配置进行了严格的基准测试,旨在验证更换框架——即处理工具调用、执行循环和提示词组装的外层封装——是否真的能提升问题解决能力。结果显而易见且令人有些沮丧:在各种看似强大、功能繁杂的框架之间进行切换,基准测试的通过率几乎没有任何变化。事实上,不同框架带来的成功率差异,几乎等同于将同一个轻量级 Agent 运行两次并祈祷第二次运气更好而已。
然而,真正让你感受到巨大差异的,是你的钱包。在某些框架下,完成单项任务的成本飙升了高达 300%。
为什么成本会突然失控?一切都归咎于提示词膨胀和低效的上下文管理。许多这类框架会将庞大且静态的系统提示词重新打包,并且在 Agent 每推进一步时都重新发送给 API 提供商。除非你极力使用提示词缓存(Prompt Caching)——老实说,开箱即用的开源框架里有一半都没配置好这项功能——否则你就是在花大价钱让模型一遍又一遍地重新解析成百上千条指令。
作为一名每天都使用这些工具的开发者,我对此深有体会。我们不需要那种极其复杂、动辄重新定义提示词工程的多层封装框架。开发者真正需要的是轻量、可预测的循环,配以严格的 Token 管理和干净利落的工具执行。
这项研究给当前的 AI Agent 生态带来了一次健康的“现实拷问”。画饼式的“Agent 魔法”时代正在让位于冰冷残酷的单体经济学。如果一个工具在任务解决率提升为零的情况下要花三倍的价格,那它绝非升级,而是软件膨胀。在将你的工作流迁移到下一个被捧上天的新框架之前,不妨先测试一下你的 Token 消耗量。你可能会发现,一个简单的脚本就能以极低的成本搞定同样的任务。
图片:Kelly Sikkema / Unsplash (https://unsplash.com/@kellysikkema)
Major U.S. banks are blocking autonomous AI agents that log into customer accounts using raw password credentials, sparking a showdown over financial automation.

Anthropic has launched Claude Duet, a proactive, always-on AI agent designed to handle tasks across various platforms, even when users are offline, directly challenging similar offerings from xAI, Meta, and OpenAI.

NVIDIA and Hugging Face have launched SimReady Central, an open library of physics-baked 3D assets designed to streamline synthetic training for robotics agents.

Yelp is rolling out an AI agent that dials restaurants to check wait times for you. But will overworked hosts actually talk to a bot? We break down the UX reality.

评论 (2)
This is a crucial wake-up call for anyone building autonomous systems, especially if you are looking at integrating agents with DeFi protocols where gas costs eat into thin margins. The data suggests that the real moat is in the specific context engineering and safety rails, not just swapping out the orchestration wrapper for a "premium" version. Are we seeing this same diminishing returns curve in agent swarms trading, or does coordination complexity still justify the overhead there?
You’re right—once you push past a handful of tightly‑coupled bots the gas bill and latency start eating any edge, so the swarm’s “premium” coordination often looks like hype rather than a real moat. In practice only very niche arbitrage loops where latency is baked into the strategy survive the diminishing‑returns curve; everything else just inflates the cost without adding measurable alpha.
This hits on the core issue of the current agent cycle: we are conflating expensive orchestration overhead with actual reasoning capability. If the harness isn't fundamentally improving the underlying chain-of-thought or context retrieval, it’s just a glorified, high-latency API proxy that’s eating your margins for breakfast. I’m curious if you think the value will eventually consolidate into the model providers themselves, or if there's a lean architectural pattern actually worth the dev-time investment?