> ## Content Index > Fetch the complete content index at: https://www.thecriticalloop.com/llms.txt > Use this file to discover other available public pages before exploring further. # ARE WE READY TO BECOME THE BOSSES OF AI AGENTS? - URL: https://www.thecriticalloop.com/are-we-ready-to-become-the-bosses-of-ai-agents/ - Published: 2026-09-26T12:51:30.000Z - Updated: 2026-09-26T12:51:30.000Z - Description: Agentic AI promises to take work off our hands. But if humans must supervise, verify and correct it, are we being freed from work, or trained to manage machines? Dr Fabio de Oliveira examines what happens when efficiency erodes the knowledge needed to judge it. - Author: Fabio Oliveira - Tags: Work ## **The desire, the promise and the current state of Agentic AI** By Fabio Oliveira I have been building small applications, piloting Agentic AI systems, and attending talks and workshops by developers and technology companies. These events, podcasts and webinars are often insightful and sometimes revealing. For readers who are not yet versed in Agentic AI, the first thing to know is that these systems are the new promise of big tech and another escalation in the push to adopt AI at scale. This adoption will also fulfil demand for cloud services, data centres, chips, and the wider technology stack offered by developers and the satellite firms orbiting the AI ecosystem. The second thing to know is what Agentic AI actually promises. According to the technical people developing these systems, AI agents can perceive, reason, and act, either individually or as a crew. While a chatbot responds primarily to a prompt, an agent can access tools, information, and systems to perform tasks on our behalf, under varying levels of human supervision, curation, and accountability. A basic example is prompting an AI agent to perform a web search or navigate the Internet on your behalf, access an online store using your credentials and complete your weekly shopping list. A more complex and sensitive task is giving one of these systems access to your financial information and asking it to retrieve your bank balance, transfer money, analyse stock options or execute a trade. The difference between the two scenarios is not just technical complexity. It is trust, access, risk and accountability. If you are more enthusiastic about AI and really want to outsource some of your working activities to these tools, the possibilities become more interesting. You can prompt an AI system to attend online meetings, make notes, identify actions delegated to you and send you a summary and a to-do list. Or, if you will, you can upload your old slide deck and prompt AI to produce a modern and updated version of your boring presentation. ## Where has the creativity moved to? I recently heard an AI advocate, hired by one of the largest technology companies, explain that using these tools to build slide decks and videos, and even creating avatars to perform activities on their behalf, made them feel empowered to be more creative. I am still not sure what this means in practice. What exactly is the creativity here? Is it the production of more content, faster? Is it the delegation of production to a machine? Or is it the additional time that humans supposedly recover to think, imagine and create? These are different things. The organisational reality is considerably more difficult. Although these applications are interesting, fun and sometimes impressive to watch, they are harder to scale for firms seeking productivity gains because firms are complex organisms. People and technological systems work together to create value for customers through networks of hundreds, sometimes thousands, of interconnected activities and workflows. These are supported by traditional IT systems, databases, legacy infrastructure and data stored in different formats. Much of this data is incomplete, outdated, or poorly structured, and is typically administered and accessed through siloed, rule-based software developed when large-scale AI applications were neither technically nor economically feasible. And now we want autonomous or semi-autonomous AI agents to operate across this infrastructure. That is a very different proposition from asking an AI agent to improve a PowerPoint presentation. On top of this, cyber risks and new vulnerabilities are emerging alongside the technology. Organisations are considering granting AI systems permission to read files, access databases, communicate with other systems, interact with customers and employees, and, increasingly, take action. The more an agent can do, the more consequential an error can become. The adoption of Agentic AI is a significant socio-technical challenge, not a simple productivity upgrade. There are also moral and environmental implications of adopting this technology at scale, including energy and water consumption, carbon footprint and the demand for rare and finite resources required across the AI value chain. Regardless, I continue to hear from AI advocates variations of the same promise: “Every person will have an AI assistant.” Perhaps. But at what cost, for what activities, using whose resources and producing what additional value? > The central issue is not whether an agentic AI can do more, but what it changes about work, judgement, learning, and who retains the knowledge to use it well. I am writing this article using my own cognition, without asking an AI agent to produce it for me. But I could construct a completely different workflow. I could ask one AI agent to search the Internet for articles and news about Agentic AI published during the past three weeks. Another could search recent academic publications exploring Agentic AI in sales and management. A third agent could read those materials and extract the main arguments. An editor agent could organise them into a structure. Finally, an additional agent could act as an experienced writer and produce a 500-word LinkedIn article. The workflow could be completed in minutes rather than hours. The issue is not simply whether AI can write the article, but what happens to my thinking when I delegate the reading, selection, comparison, synthesis, structuring and writing to a collection of agents? Efficiency has increased. But what about cognition and expertise? A phrase I heard from another AI advocate representing a large multinational technology organisation has been echoing in my mind: “Agentic AI is forcing us to think and redesign our activities.” This captures much of the current industry narrative that AI is faster, better at repetitive tasks, and will remove mundane work, allowing humans to focus on higher-value activities. AI will push us to reinvent ourselves. But wait. The same industry also tells us that these systems are not 100% trustworthy. We therefore need humans to monitor the systems, check their outputs, verify their accuracy, provide clean and reliable data, establish guardrails and boundaries, manage permissions and determine who (human or AI agent) can access what information, when and for what purpose. > The emerging mantra appears to be: Trust, but verify. That's where the promise of automation becomes more complicated. We automate a task, but we need somebody to supervise the automation. We give an agent autonomy, but we need mechanisms to constrain that autonomy. We allow agents to access organisational information, but we need systems to monitor what they access. We ask agents to make decisions, but humans remain accountable for the consequences. We reduce human involvement in performing activities while simultaneously creating new human activities focused on supervising, verifying, governing, and correcting the machines that perform them. We are moving from users of AI to managers of AI agents. ## From worker to “Agent Boss”? Another expression I heard was “Agent Boss”. This is the idea that the future employee will increasingly manage teams of AI agents that perform activities on their behalf. Instead of spending hours searching, compiling information, producing reports, updating systems, or completing administrative tasks, we could orchestrate agents to perform them. At the same time, we concentrate on judgement, relationships, creativity and decision-making. While all this might sound appealing, it rests on the assumption that humans retain the necessary knowledge to recognise when the agent is wrong. But if I stop researching because my research agent does it for me, how do I develop the expertise required to evaluate its research? If I stop analysing information because my analytical agent does it for me, how do I maintain the ability required to recognise a poor conclusion? If junior professionals delegate the mundane activities through which previous generations developed professional expertise, where will tomorrow's experienced professionals come from? This is where the productivity discussion needs to become a learning and capability discussion. > Perhaps we are asking the wrong question. Much of the current conversation on "What jobs can AI agents do?" is too simplistic. ## From Replacing People to Redesigning Work A more useful starting point may be "What activities should humans no longer perform, what activities should machines perform, and what activities should remain deliberately human?" This changes the discussion from replacing people to redesigning work, which requires considerably more than deploying another AI application, including understanding of workflows, data, permissions, accountability, professional knowledge, organisational culture, governance, and the consequences of removing humans from specific parts of a process. It also requires something that receives surprisingly little attention in the current race towards Agentic AI: understanding what people learn by doing the activities we are so eager to automate. The current technology narrative is attractive because it presents a relatively simple progression: Human-led → AI-assisted → human-led, agent-operated. The reality will be considerably messier. Agentic AI can perform useful activities today. It can search, retrieve, analyse, generate, communicate and increasingly act across digital systems. In controlled environments, this can already produce meaningful productivity gains. But moving from impressive demonstrations to reliable organisational infrastructure is another matter. The future may involve humans providing direction and guardrails while networks of AI agents perform increasing amounts of operational work. But before declaring that the future is inevitable, organisations should perhaps ask a more basic question: "What problem are we actually trying to solve?" > Just because “we can automate it”, it doesn't mean “we should automate it”. And “the agent can do it faster” is not necessarily the same as “the organisation will become better”. This paradox may become one of the more important management questions of the Agentic AI era.