Dr. Neil's Notes
Software > Development > Being Early
Agents Before the Agent Era
Introduction
In 1997, the same year we were building WebPrecinct, another idea was taking shape: Agent Task Force. The concept was intelligent agents that would sit on your desktop as personas, helping you manage daily tasks. The agents would be driven by the recognition that neural networks were moving toward smarter, more adaptive ways of working.
The world would eventually catch up with this idea. It took approximately twenty-five years.
The Idea
The core insight behind Agent Task Force was straightforward: software should adapt to people, not the other way around. Instead of learning menus, shortcuts, and application workflows, a person should be able to describe what they want in plain terms, and an agent should figure out how to get it done.
This was 1997. The computing infrastructure needed to make that work at scale did not yet exist. Consumer hardware was a fraction of what it would eventually become. The machine learning techniques available were limited compared to what would emerge over the following two decades. The idea was right. The timing was off by a generation.
What Made It Hard Then
The gap between the concept of intelligent agents in 1997 and the reality of what could be built was enormous. Neural networks existed and were understood in principle, but training anything sophisticated required compute resources that were inaccessible outside of research institutions. The data required to make agents genuinely useful was not yet available in the structured, accessible forms that would later make modern AI systems possible.
Building a convincing agent persona in 1997 meant writing a great deal of rule-based logic and accepting significant limitations in what the agent could handle. The further a user strayed from anticipated patterns, the more visibly the system broke down.
The ambition was ahead of the infrastructure by a wide margin.
Twenty-Five Years Later
By the early 2020s, the world had arrived at the destination Agent Task Force had pointed toward. Large language models made conversational, adaptive, task-oriented agents genuinely possible. The compute was available. The data was abundant. The tooling had matured over decades of incremental progress.
The agents that now sit in every browser, operating system, and productivity tool are, in essence, the idea from 1997, finally running on the infrastructure it always needed.
The lesson is not that the 1997 work was wasted. The people who built early agent systems developed intuitions about human-computer interaction, task decomposition, and failure modes that shaped the field for decades. The ideas filtered forward. The patterns survived.
Being twenty-five years early is still being right.
Part of the Being Early series.