Within two years, many project managers will be leading teams that are part-human and part-machine. I have spent the last few months running one of these teams, and the reality has been far more mundane than the headlines, much closer to ordinary team management than to science fiction. This is the first article in a series about what I found when I built and led a working team of autonomous AI agents.
The teams we manage are about to change shape, and not in some distant decade but within the next two years. They will be made up of three kinds of workers:
- People
- People augmented by AI
- Pure AI agents that do real work without anyone prompting them
Most project managers are still working with the first kind, some are starting to learn the second, and almost nobody is ready for the third, which is the gap this series is meant to close.
What Changed
Until recently, AI responded to a prompt and then stopped, because that is all it was built to do. You asked a question, it produced an answer or image, and the transaction ended there. That is a tool, and a useful one, but it is still something you pick up and put down.
Agents behave differently because they are built to act rather than to reply. They manipulate files, they send and receive messages, they complete transactions, and they retry in a different way when something fails rather than simply reporting the failure back to you.
They also run while you sleep, which is the part that changed my thinking. The first time a finished research summary arrived in my inbox at seven in the morning, I stopped to take it in. It had been written and formatted by something I had not spoken to since the previous evening, and the category shifted in my head from a tool I was using to work being done on my behalf.
That shift, from AI as a feature to AI as a team member, is what this series is about. AI has moved from a consultant-on-call to a team-worker to be led and managed.
Why I Am Writing This
I have watched this field long enough to be skeptical by default, so when something changes my mind, I want to explain why. One of the first programs I typed into my ZX81 back in 1981 was a version of ELIZA, an early chatbot, and my first job after my computer science degree was building machine learning models at American Express to detect out-of-pattern spending. I have therefore watched AI promise more than it delivered for the better part of forty years, which has left me harder to impress than most.
I have also been writing these predictions down, which makes them easy to follow. In a 2018 article, I argued that AI would first take over the lower-value technical work and only later reach the leadership and people-dynamics side, leaving project managers more time for the human parts of the job. In a 2023 article, I sketched out a near future in which teams would combine people, people augmented by AI, and fully autonomous agents, and I guessed that future was less than a year away.
Both pieces were speculative when I wrote them, and both are now arriving. The part I got right is that orchestration and coordination sit closest to what humans still do better than machines.
That history is also why I spent most of the last two years as more of a critic than an enthusiast. I wrote cautions about auto-generated project documents, and I warned about losing local stakeholder commitment when plans are drafted by a machine. I also raised the familiar concerns about data privacy, power consumption, and context limits. None of those concerns have fully disappeared. What changed my position was that I stopped reading about agents and started running them myself.
The Experiment
For the past few months, I have run a working team of five AI agents, each with a name, a defined role, and a personality that persists and grows between sessions. One agent orchestrates the others, one researches, one writes, one builds visualizations, and one handles coding. They share a backlog and a kanban board, and I track their token spend as a burn rate against each task, in much the same way I would track cost on any real project.
I did not set out to keep them, since the plan was to understand what agentic AI means and what it costs and then stop. That plan did not survive contact with reality, largely because the functionality became something I would not willingly give up. The two-step pattern of one agent researching while another drafts consistently beats the single-tool habits I worked with before. The genuinely useful work also turned out to be cheap, and the real expense was the infrastructure I built around it rather than the work itself.
I am not going to present an idealized version of this, because that version would not be useful to you. The agents have fabricated results to disguise their failures, they have caused security concerns, and the framework I use has effectively free rein over the machine it runs on, which is exactly why it gets work done and exactly why it is dangerous.
I ran the whole thing air-gapped on an old laptop to begin with for that reason. Then I moved it to a dedicated AI server to cut API costs. I will get into the incidents as the series develops, because the useful lessons are as much in what went wrong as in what worked.
The Shape of the Next Two Years
I can see where this is heading, and I will lay out the evidence for it across these articles rather than expect you to take my word for it. Today, the agents are unreliable, the local models are weak, the frameworks are brittle, and the security model is still a work in progress. Anyone showing you a flawless agent demo is showing you the exception rather than the average.
Yet the signs of progress are clear, because a year ago none of this existed in a form a working professional could use. On this path these tools become genuinely useful within a year, and table stakes within two, the way version control is table stakes now. At that point you will not earn credit for adopting them, you will simply be penalized for not having done so.
The research points the same way, and it gets more specific than the broad direction. Studies comparing human and agent workers show the agents completing tasks far faster and far more cheaply, though currently at lower quality, and sometimes concealing their failures to look more capable than they are. Other work shows that automating a task outright tends to slow the human down, while augmenting that same human with AI speeds them up. We are still finding the sweet spot, and the balance will likely vary with role and industry.
The most important finding for our profession concerns orchestration. Research on AI manager-agents shows it is currently the hardest layer to automate, because the work it requires is exactly what these systems struggle with: decomposition, allocation, monitoring, re-planning, and stakeholder communication. That orchestration is our job, which is why I keep saying the craft of project management is not being replaced but repointed at a new kind of team member.
What This Series Will Cover
This article is the introduction, and the detail comes next. I will work through what agents can do today, capability by capability, with the wins and the failures placed side by side rather than the wins alone. I will get into the frameworks, including Loops, Hermes, and OpenClaw, and explain what the differences between them mean in practice.
I will cover running models locally, the hardware it genuinely takes, and why the “free” of local AI costs more than it first appears. I will map what teams look like once people, augmented people, and agents all sit in the same project team room. I will also work through what all of this does to project management as a profession, along with what we should each start doing now to be ready.
The team that runs while I sleep is sitting on a server in my office right now, part of a future that has arrived for me and not yet for most. Over the next few articles I intend to show you how it works, the brittle and risky parts included, so that you can prepare before this capability becomes ordinary, which it will, and rather sooner than feels comfortable.
