Large Language Models (LLMs) like ChatGPT and Claude have yet to match the versatility of human workers. This is partly because AI relies on uploaded data for context. Thus, AI tools have primarily served as co-pilots, helping users complete specific tasks, but unable to assist autonomously.
Beyond Co-Pilot Assistance
Last month, Anthropic released a new function via its API – Claude ‘Computer Use’. Despite its innocuous title, Computer Use represents the closest any mainstream AI has come to human-like agency.
Anthropic’s Beta Computer Use enables Claude to interact directly with software environments and applications – navigating menus, typing, clicking, and executing complex, multi-step processes independently.
This functionality mimics robotic process automation (RPA) in performing repetitive tasks, but it goes further by simulating human thought processes, not just actions. Unlike RPA systems that rely on pre-programmed steps, Claude can interpret visual inputs (like screenshots), reason about them, and decide on the best course of action.
For instance, a business might task Claude with organizing customer data from a CRM, correlating it with financial data, and then crafting personalized WhatsApp messages – all without human intervention. A developer might request Claude to set up a Kubernetes cluster, integrating it with the right configurations and data. Such capabilities make it feasible to delegate work to Claude in the same way one would assign tasks to a junior employee.
However, there are trade-offs: relying solely on Claude’s Computer Use can be slow because it mimics human actions step by step. Furthermore, Computer Use as stated in the name needs to have exclusive access to a computer when working.
The Value of Multi-Agent Configurations
In my article last month on AI Agents: Are We Ready For Machines That Make Decisions? I explored the controversial question of AI agency. Tools like Computer Use don’t offer true autonomy but simulate it effectively, creating opportunities for business innovation. This month, I spoke with Daniel Vassilev, CEO of Relevance – a platform that provides AI agents – which rely on a deeper technical integration than Anthropic’s Computer Use – about the practical applications of these technologies.
“Agents let teams unleash their output based on their ideas, not their size,” Vassilev explains. Each set of agents provided by Relevance is estimated to handle workflows equivalent to what would typically require five full-time employees (FTEs). This could include activities such as lead qualification, personalized onboarding, and proactive customer success outreach—tasks that would be prohibitively resource-intensive without automation.
While automating single workflows is beneficial, the real value lies in deploying multiple specialized agents. Just as businesses organize teams by expertise, AI agents designed for specific tasks—like research, outreach, or documentation—can collaborate to drive exponential productivity. These agents integrate seamlessly across workflows, compounding efficiency gains without interpersonal friction or the need for additional human oversight.
The Autonomous Edge
The key distinction between co-pilots and autonomous agents lies in execution. What sets autonomous agents apart from co-pilots is their ability to execute tasks independently. As Vassilev puts it:
“A co-pilot makes you twice as productive, but an autonomous agent lets you delegate the work entirely, leaving you to review the output.”
As an example, Relevance uses their own AI agents to; research new customer signups in order to generate tailored recommendations, onboard users by pre-creating tools customized to their needs, and follow up with personalized communications. These agents shift human roles from task execution to oversight, freeing up time for strategic and creative work.
Trust and Guardrails
Despite their potential, AI agents are not infallible. Vassilev likens deploying AI agents to onboarding a new hire:
“You wouldn’t let a new hire send an email to your customer’s CEO without oversight. Similarly, AI agents require a strong human-in-the-loop process.”
The need for ensuring that AI agents are performing safely is reliant on setting guardrails about what they can and can not do and ensuring that they are trained properly – similarly to a junior employee.
Challenges and the Path Forward
Despite their promise, autonomous AI agents face hurdles. As Vassilev notes, many automation projects fail not due to technical shortcomings but because of organizational wisdom gaps:
“Unique processes often reside in the minds of subject-matter experts, making them difficult to document and automate.”
However, combining Anthropic’s Computer Use with multipe AI agents opens up automation possibilities that would have been inconceivable even 6 months ago for non-repetitive, creative, or low-scale activities.
As tools like Anthropic’s Computer Use (it is still in Beta) and Relevance’s AI agents mature, the potential for businesses to achieve more with fewer resources will expand. Organizations will no longer be constrained by headcount, human roles will shift toward oversight and innovation, and ambitious goals and innovative solutions can be unlocked. Exciting times.
