Before the AI Employee

In April 1994, Wired described General Magic's Telescript. A message sent through the network would no longer be a passive packet. It could become a travelling program, carrying its sender's instructions into remote systems and performing work there.

The designers imagined agents arranging travel, retrieving prices, or carrying medical information. They also gave the agents permits, identities, resource limits, and technical restraints. The earliest commercial vision already contained questions that returned thirty years later: who sent it, where may it travel, what may it use, and who remains accountable for its actions?

General Magic was building infrastructure for representatives of users. The agent remained outside the organization, dispatched into a network to complete an errand.

Researchers associated with the MIT Media Lab incorporated Agents Inc. in 1995. Its system learned musical preferences and recommended artists by comparing one listener with others. The company later became Firefly Network, partly because the word agent sounded threatening to the people expected to trust it.

Firefly represented a quieter form of agency. The system observed preference, built a model, and made a recommendation. The commercial principle was significant: a program could hold an imperfect understanding of a person and use it to act on that person's behalf.

The agent had entered the product. It had not entered the organization chart.

The direct ancestor of the modern AI employee emerged from back-office automation. Blue Prism was established in 2001 and developed software robots for repetitive business processes. Its first commercial version arrived in 2008.

Blue Prism called the system a Virtual Workforce. The robots could be scheduled, supervised, and audited through a central control room. Business specialists could define processes without building a new application for every task.

By 2016, Blue Prism described a large British utility operating more than 300 robots under the supervision of two people.

The robots followed processes defined in advance. They could branch, recover, and scale, but they could not interpret an open-ended responsibility in ordinary language. They were a disciplined industrial system, closer to machinery than colleagues.

In 2014, Deep Knowledge Ventures announced that an analysis program called VITAL had been appointed to its board. The system examined life-science companies and made recommendations. The appointment attracted attention because a program was being described as a participant in corporate judgment rather than as a back-office tool.

The legal and organizational reality remained human. VITAL was an analytical system rather than a corporate person capable of carrying formal duties. The title arrived before the autonomy, a pattern that would return whenever companies gave software executive names and offices.

The DAO, launched on Ethereum in 2016, tried another route. Instead of placing automation inside a conventional hierarchy, it encoded parts of the organization itself. Participants considered proposals, voted, and relied on smart contracts to carry out approved decisions.

A flaw in the code was exploited and the Ethereum community intervened. The episode established a principle that remains central to agentic companies: autonomy increases the reach of both competence and error. Governance cannot be added after deployment as decoration.

Large language models changed the unit of automation. Earlier systems usually began with a documented process. A language-model agent could begin with a role, an objective, a set of tools, and a description of acceptable behavior.

ChatDev arranged specialized agents around design, coding, and testing. MetaGPT assigned roles within an assembly-line model and encoded standard operating procedures into prompt sequences. These were research systems rather than operating companies, but they established the architecture of the synthetic office.

A company could now be represented as a set of roles passing work to one another in natural language. An organizational brief had become an executable interface.

In 2025, Wade Foster disclosed that Zapier was using more than 800 internal AI agents while the company had fewer than 800 people. Examples included meeting preparation, message drafting, research, and reporting.

The ratio was widely repeated. It was also contested because many examples resembled narrow workflows rather than artificial colleagues with broad responsibility. The criticism revealed that the word agent now covered systems with very different levels of autonomy.

Zapier's more durable lesson concerns the division between uncertain and stable work. An agent can investigate an ambiguous problem and help discover a process. Once the process stabilizes, ordinary automation can execute the repeated steps while the agent handles exceptions.

Language models interpret. Deterministic systems repeat. Humans retain judgment where consequences are substantial.

By 2026, companies were publishing synthetic organization charts. Founder Coach described 25 agents across finance, marketing, sales, strategy, and relationship management. An AI Chief of Staff collected reports from the others. The arrangement reproduced familiar coordination problems: the central agent became overwhelmed, and subagents once removed 571 files while attempting to reverse a project.

Creem used a distributed structure. Publicly described agents were paired with particular people and assigned continuing domains across go-to-market work, support, infrastructure, engineering, and finance. They met in Slack at 3 a.m., exchanged techniques, and published reusable skills. Each agent had a human counterpart and a recognizable place in the company.

ClickUp reported 3,258 agents working alongside 1,300 people. Teams across the company created the agents themselves. This was an agent mesh rather than a compact synthetic department. The central challenge shifted from adoption to governance: ownership, permissions, duplication, maintenance, evaluation, and retirement.

Glitter AI showed the founder-controlled model. Its solo founder described agents that organized customer feedback, drafted support replies for approval, and operated a content-production pipeline. Customer communication remained behind a human boundary while software compressed the repetitive parts of the work.

These arrangements shared a technology but represented different institutions. Founder Coach built a hierarchy. Creem built counterparts. ClickUp built a mesh. Glitter built production lines around one founder. Their agent counts were not directly comparable.