Why we decided to open source Relate
We have been deep in the weeds of various clients' workflows. Each one could
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
We have been deep in the weeds of various clients' workflows. Each one could have been its own SaaS tool. Our view is that going forward this will no longer be the case. Agents can now produce decent user interfaces with hardly any guidance, and pretty good ones in the hands of someone with "taste".
So does this mean that software is dead? Not in our view. There are many low-level components which we should not be rebuilding for each app. Yes, agents do know how to implement an idempotency key. They also know when to implement one. But they will not implement one each time they should. The knowledge exists in the model weights, but that knowledge may not be applied. In production, you will hit a bug and need to fix it. There are millions of similar cases.
There is a gap between what a model knows and what it reliably does. Researchers have documented a related “knowing-doing gap”1The knowing-doing gap in LLM tool use in tool use: a model can recognise that a tool is needed and still fail to use it. Knowing the right thing to do does not guarantee doing it when it matters.
So why go through the pain of "hand-rolling" low-level components from scratch? We believe there is great value in using software that is battle-tested across multiple deployments. Especially when it is open source and can be patched by developers across the world (or their agents).
Defining a permission-aware business ontology, which connects your systems, is exactly this sort of problem. Lots of boring, repetitive code that should work efficiently and without errors. That is Relate.
You define your business once: customers, invoices, campaigns, and how they relate.
We are early, not ready for production workloads, and definitely not yet battle-tested.
We would very much appreciate your feedback or a GitHub ⭐.
Cheng, Y. et al. (2026). Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use. Studies the gap between recognising when a tool is needed and actually invoking it.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。