开发者分享实战心得：为什么该自己搭agent框架而不是用别人的黑盒，还点名了LangChain DeepAgents和Claude Code的用法对比。
Matt Stockton在帖子中讨论了AI agent harness的透明度问题，认为私有harness增加了非确定性放大效应。他推荐在生产环境中使用LangChain DeepAgents作为轻量级harness，并指出自定义harness在控制、成本和模型灵活性上的优势。他本人用Claude Code和Codex做交互式编码，并开始用Pi构建自定义harness。他预测企业将加大对自定义harness的需求，成本是主要推动因素。
good post on the value of knowing whats going on in your harness "For agents within an actual produ...
good post on the value of knowing whats going on in your harness "For agents within an actual product, I have gravitated to using LangChain DeepAgents" Matt Stockton @mstockton The frontier model lab harnesses are amazing - don’t get me wrong. I love Claude Code and Codex, and the lab companies have the unique moat where they can train their models such that they perform better in their harnesses. But I think the ‘opaqueness’ of what a harness is truly doing - is becoming more problematic over time. LLMs are non-deterministic from the start - and adding an opaque level of abstraction (eg a private harness) on top of the LLM is effectively a ‘non-determinism amplification function’ - you are making something that is already hard to direct - potentially even harder to direct, without truly being able to inspect how it’s being directed. Sure - agents markdowns, skills, custom tools, the ‘LLM wiki pattern’ - all this context engineering helps to constrain the behavior - but there is still an opaque intermediary It might still be too early for companies to really invest a bunch of time in harness engineering, but I do think for non-experimental, production grade stuff, ‘build your own purpose-built harness’ is the future. Not only for control and correctness purposes, but also for cost and model flexibility purposes. How has this practically affected my work? - I still use Claude Code or Codex interactively for almost all of my code building. - I use claude -p for a lot of background jobs as well via CI/CD. - I have begun experimenting with custom Pi-based harnesses around my knowledge wiki and also for specific code projects. Once a project is more mature, adapting the context from Claude Code into a more purpose built harness in Pi has started to show signs of value. - For agents within an actual product, I have gravitated to using LangChain DeepAgents. I feel like their base harness is less heavy, and their abstraction for building agents that use skills and tools is really nice. Anyway, I don’t know where this is all going, but I’d make a bet that there will be an increased demand and interest for companies to built custom harnesses. It does seem like cost is going to be be the issue that starts the snowball rolling down the hill here 🔗 View Quoted Tweet 💬 3 🔄 0 ❤️ 8 👀 1492 📊 4 ⚡