An agent that demos well takes two or three weeks. An agent that frontline staff use daily, and that a manager will sign off on, takes three to six months. The distance is not model capability; it is data preparation, knowledge structuring, workflow embedding and permission alignment.
Received wisdom says choosing the right model matters most. Our own time tracking says the opposite: model selection is about twenty percent; eighty percent goes to data preparation, knowledge structuring, workflow embedding and permission alignment.
From question answering to autonomous execution, capability rises and controllability falls; the gap between them is the risk exposure. The more an agent can do, the more it can do wrong.
Models depreciate continuously. Only three things accumulate and appreciate:
Structured, traceable domain data and knowledge that AI can consume directly.
Tool calling, memory, permission sandbox, evaluation loop. The model is a swappable engine; the harness is the chassis that accumulates.
Repeatable expert workflows packaged as versioned skills. The growth curve of the skill library is the learning curve of the system.
A core team member is second author on the paper.
Semantic embedding methods and clustering applications for large-scale enterprise data.
An interpretive framework for Transformers from the perspective of filtering theory.
Multilingual enterprise entity alignment and clustering.
Filed with the Cyberspace Administration of China: 网信算备440309441584901250013号.
Rust-layer rebuild; 48 Skill and Harness classes.
Where our positions sit relative to the field.
Joint postgraduate training with several universities, with company mentors throughout, turning delivery methodology into transferable organisational capability.
Tell us which scenario is still not in production: where the data lives, who uses it, and who signs off. One workshop gives you a verifiable path forward.
