It is very annoying, on a personal level, when karma comes around to bite your ankles. A long time ago, I was happy at work, in a chaotic way. I developed things – from hardware and software to courses and programmes; I wrote texts, built things, and generally messed around. Then, more than twenty years ago, an improbable sequence of events very suddenly made me the head of the academic department where I worked, comprising twenty-odd colourful personalities including my previous (tenured) boss.
I tried my best, but the situation was non-trivially challenging, and by and large the experience was not a happy one for anyone within the blast radius. Leadership is something you need to learn and practise – most people do not have leadership or even management skills installed and configured at birth. I certainly did not, and some years later I drew the only possible conclusion and left. I promised myself never to try to manage other people again, and also never to take part in a development discussion in any role whatsoever.
I have never regretted this – I went back to developing things: courses and programmes, reports and financial organisations, and basically enjoying the things I do. There’s a very nice colleague who keeps scheduling the occasional development discussion, but I prefer to think of them as a pleasant chat with a friend over lunch (except someone always seems to forget the food part!). It has all worked out very well. And then karma, in the unlikely guise of generative AI models, came around.
The rise of agentic AI
The early GenAI models, vintage 2023 and 2024 – ChatGPT and friends – were fascinating and impressive in many ways, but they were not quite there yet in terms of utility and productivity. They were tools for finding and summarising information, for generating text and other forms of media; entertaining to play with, frustrating in their non-repeatable behaviour and frequent hallucinations. The potential was clearly enormous, but the actual value added by more elaborate B2B spam emails and artificially generated pictures of cute squirrels was simply not there.
This is no longer true. There are now at least two interlinked areas where the productivity impact is already significant in an order-of-magnitude, moving-the-decimal-point kind of way: discovery and development, and software and automation. The underlying LLMs have developed significantly, but the game changer is that they are now agentic – they can use tools, and it (mostly, increasingly) works.
Discovery and development:
There are cases galore, but two will suffice to make the point. The Harvard + Procter & Gamble experiment from the summer of 2024 persuasively showed that, when developing ideas for new products, teams consisting of experts and AIs consistently outperformed teams with only flesh-and-blood experts. It wasn’t even close, even with that generation of not-yet-agentic AI technology. And then there is Google’s AlphaEvolve
– an AI capable of systematically developing and refining tools and techniques (code) to solve hard problems. The trial run made short work of dozens of “best” solutions to problems in applied mathematics, and again, it was not even close in many cases. This is moving fast, but it is already clear that R&D should be done by human experts together with agentic AI models if it is to be competitive. It’s not even close.
Software and automation:
Generative AI models, being software themselves, are particularly well suited for writing code. This is partly because programming languages, frameworks and libraries are well documented compared with almost anything else, and partly because the tools used by programmers – compilers, linters and so on – are easy for agentic AIs to use. They can iterate like coders do – write code, review it, fix errors, compile it, debug it, run it, test it, fix it – all on their own until it works. Using a modern agentic coding AI interface like Cursor, Codex or Claude Code with the latest underlying models (GPT-5, Claude 4.1, Gemini 2.5, at the time of writing) is extremely impressive. “Vibe coding” has tainted the perception of GenAI coding a little – if you do not know what you are doing you will not necessarily get the tools to produce very good code, surprise, surprise – but in the hands of even a moderately experienced software developer, and managed in a disciplined way, these tools really boost productivity quite significantly.
These two fields are currently the clearest examples of major domains where already, if you do not use current agentic AIs, you are basically obsolete. There are others, like analytics, and more are coming. In order to deal with this we need to forget about the clunky Copilots of 2024 and take a radically different approach.
Managing AI, not just using it
With agentic AI – generative AI models that themselves use tools – it is no longer useful to just treat them as tools that we use. We need to think about them as team members to whom we delegate, and that means they need to be managed – supervised, orchestrated, told what to do now and what to do next. We do not need leadership – these things are not driven by emotions – but we most certainly do need management. A tool is something we use interactively. A team member is someone, or these days also something, who can accomplish things autonomously. An interactive chatbot is a tool. A deep research agent or a software development agent like Claude Code, capable of operating independently, iterating and using tools as needed for hours, is best seen as a part of your team.
In certain domains I no longer do things in the sense of implementing or executing, but instead in the sense of planning, managing, and teaching. I write very little code, but I spend the time working out how to specify clearly what I want the code to do, and how to sequence the writing of the code into well-defined steps for an agentic coding AI. Often I use another AI to help me with this, and with developing and running the unit tests. The outcome is that more and better code gets written.
When I need to learn something, or write something, I will task a deep research agent with reporting on the state of the art, and then talk with it to clarify and investigate things that interest me. These things take place concurrently, and require more, or less, or almost no active supervision by me, depending on the task. Sometimes the team is idle, sometimes several members of it are busy round the clock, monitored by another AI.
The ancient (2023–24) art of prompt engineering has evolved into the discipline of context management. For your AI agents to accomplish what you need them to do you must make sure they see exactly the information they need to see. To the internal system prompts of LLMs we add top-level background prompts, then project- and task-specific prompts, examples and structures, documentation and specific tools (often MCP servers and services) and resources.
It’s a core part of management – giving your team the context, directions and resources they need to perform. You need to know the strengths and weaknesses of your team members and allocate the work accordingly. Then you need to do project management to get your team to do the necessary steps in the correct order. And then you need to make sure your team members don’t go following black-on-black instructions from dubious servers.
The HR (or should that be AIR?) grind is always a component of management. You need to review performance and cost, figure out who’s good at what, look for new skills and talents, hire and fire. Goodbye Perplexity – maybe apply to rejoin the team when you again have a distinctive advantage? Hello DeepSeek – are there some tedious and clearly not sensitive tasks you could do cheaply, freeing up the more expensive team members? Dear Midjourney and Leonardo – I need to fire one of you, please fight it out!
Should I pay you piecework rates (tokens) or have you on retainer (subscription)? How do I keep you trained (contexts updated) as time goes by and processes evolve? Management responsibilities just keep piling up.
We are all little managers now
A flesh-and-blood white-collar worker who cannot manage and direct a team of agentic AI assistants is, in many cases, no longer baseline productive. This is what the “AI first” trend means – there is no point hiring or retaining workers who do not bring, adopt, and manage a supporting AI team. Of course there is enormous inertia – obsolescence takes time. Maybe you can make it to retirement without having to manage your own team of AIs? Well, maybe – I’m 58, writing this, and I did not.
Scary? Not when you realise that your team is there to leverage your competence and productivity. Without you, they are nothing. With them, you can get more and better things done. You just need to know how. We are training and learning against a rapidly moving target, but that does not mean it’s too late or too early to start.
We need organisational support, new structures. There is massive upskilling needed, at all levels, as everyone will need to become proficient and stay current in how to manage teams of AI assistants, and how to operate productively in teams consisting of very different kinds of intelligences. The current crop of “Copilot-for-Excel-for-dummies-by-idiots” courses are a start, I suppose, but this whole field needs a much more fundamental approach. Let’s talk about that!
And I’m a manager again. Good grief.