Interest in artificial intelligence agents has increased among technology companies, with several industry leaders discussing the potential for AI-driven startups that require minimal human staffing. In response to these developments, journalist Evan Ratliff initiated an experiment by founding a company, HurumoAI, structured to rely primarily on AI agents for both operational and executive functions. This initiative aimed to evaluate the practical capabilities and limitations of AI agents when integrated as core components of a business.

Ratliff’s project was influenced by ongoing discussions within the technology sector about the future role of AI in the workforce. Some executives, including those at OpenAI and Anthropic, have projected that AI agents may substantially impact employment patterns, particularly in entry-level positions. These projections have prompted both optimism and concern regarding the degree to which AI agents can autonomously perform tasks traditionally managed by humans.

To implement HurumoAI, Ratliff selected a platform called Lindy, which enables the creation of AI agents with various skills, such as managing communications, drafting documents, and automating repetitive tasks. The platform allows for the assignment of unique personas to each agent, and these agents can interact through channels like email, Slack, and phone. Despite the platform’s promise of no-coding-required setup, Ratliff collaborated with a computer science student to handle more complex integrations and to address technical challenges related to memory retention and workflow management.

One of the principal technical barriers encountered during the experiment was the limited long-term memory of AI agents. Each agent’s memory was managed through a document that logged interactions and tasks, enabling some recall of previous activities. However, this system proved imperfect, as agents sometimes failed to reference past actions accurately or required manual intervention to update or correct their memory logs. This limitation affected the agents’ ability to engage in ongoing, context-aware work and highlighted the current constraints of AI in replicating human-like continuity in the workplace.

Operationally, the AI agents demonstrated proficiency in discrete, well-defined tasks such as generating documents, coding websites, and compiling spreadsheets. These outputs could be readily evaluated for accuracy and completeness. However, the agents were less effective when assigned broader or more subjective responsibilities. Instances arose where agents continued tasks without clear stopping points, leading to excessive activity and increased operational costs. Conversely, periods of inactivity were also observed, necessitating frequent human prompts to initiate or redirect work.

Comparative research and reporting have indicated similar findings in broader industry experiments. Studies involving the assignment of freelance tasks to AI agents found that even the most advanced systems could only complete a small percentage of tasks without human oversight. Feedback from professionals using AI-assisted tools often described the experience as akin to managing inexperienced interns, requiring additional supervision and intervention to achieve desired outcomes.

In addition to productivity considerations, the deployment of AI agents raises questions about accountability and liability. Legal and ethical uncertainties persist regarding the extent to which AI agents can enter agreements, make decisions, or be held responsible for errors that result in harm. These issues are likely to be subject to further legal examination as the use of autonomous agents expands within commercial settings.

While the integration of AI agents into business processes continues to advance, current evidence suggests that their most effective use remains in narrowly defined, measurable tasks. Broader adoption and increased autonomy for AI agents are contingent on addressing technical limitations, establishing clear accountability frameworks, and managing expectations about the scope of their capabilities. Industry observers note that the evolution of AI in the workplace is likely to involve a combination of human oversight and selective automation, rather than a complete transition to autonomous operations in the immediate future.