The Floppy Disk Era of AI Is Ending

I'm old enough to remember when the 3.5 inch floppy disk became mainstream. Until then we'd been using the 5.25 inch disks, which were cumbersome and felt fragile - one bend, one stray magnet, and a week's work was gone. When the new disks arrived in their rigid shells and bright colours, it signalled a change in the way information moved around. The folklore at the time was that they'd been designed to fit in a shirt pocket. The design goal was literally that you could slip your work into your pocket and carry it around.

And carry it we did. Before the internet reached the office, the way a file got from my computer to yours was that I saved it to disk, ejected it, and walked it down the hall. There was even a name for this: the sneakernet. For the record, I would have been rocking my Adidas Romes around that time, and I was more likely in high school than in an office.

I've been thinking about the sneakernet a lot lately, because it describes what working with AI has felt like until now. When one of my AI agents drafts something I have to metaphorically carry it to the next one. I need to copy, paste and re-explain the context every time. Another AI agent might produce research that I then ferry into a document, summarise for a colleague, who in turn feeds it to their own AI, which probably asks a question mine already answered. The models have become astonishingly capable, but until now we're still walking the disk across the room.

The story so far

My colleague Frédéric Etiemble tells a story he calls the AI agentic revolution. It describes the old world (the world before ChatGPT, or "B.C.") when you had a task to do, you figured it out, asked someone in your network, or Googled it. Then in 2022 the AI assistants arrived, and for a lot of knowledge work you could simply ask. Around 2024 came the world of AI tools: tens of thousands of them, one for every job, with you still holding the controls. In 2025 the agents emerged - systems you hand a goal to, which then act on your behalf. And what's emerging now is the world of connected agents: teams of them, coordinating on bigger goals, with something like an AI chief of staff at the centre.

Fred anchors the story with a William Gibson line that has never felt more accurate: the future is already here, it's just not evenly distributed. In my work with teams navigating AI, I see the whole arc alive at once. Most organisations sit somewhere between assistants and tools. A few individuals, usually quietly, are running agents. Almost nobody has reached connected agents, and the ones who have will tell you, honestly, that it's held together with sticky tape and string.

Here's the thing I want to add to Fred's map. Every stage of that story, up to and including agents, is still single-player mode. Your assistant lives in your chat history. Your tools take your input and hand you back output. Even a genuinely autonomous agent is your agent, running in your session, with your context. The payoff is real, which is why so many of us carry the load - but it has been a solo game. The moment you try to make it a team sport, everything gets harder.

The network is being built

That's starting to change, and I have some skin in this game, because one of the emerging answers is part of my daily routine.

I've long been a proponent of Getting Things Done, with its single inbox and its philosophy of "mind like water" - the idea that you shouldn't be carrying cognitive load around in your head. For me, that philosophy now interfaces with Nate B Jones' Open Engine pattern: work goes into a queue that both humans and agents can read, and the ticket, not the chat, is where everyone meets.

In practice, whenever I have a thought, I talk to my Hermes agent by voice, through my Apple Watch. It's effectively my "Chief of Staff": I ask it to shape the thought into a task and assign it to the appropriate queue and the appropriate agent for the kind of work at hand.

I have different agents on multiple computers each geared for different strengths such as coding, or working with local project files etc. A key point is that this network of agents is diverse – the models from different providers, each wrapped in different harnesses, and with different cost and risk profiles. This is different from being beholden to one product, or putting all work through the same pipeline, and I think it is more reflective of how organisations will need to manage work.

My queue is literally a Kanban board in Linear (similar to Jira or Trello). So a task lands, and then gets picked up automatically by the assigned agent. Later I come back to find the completed work waiting: flagged for my input where judgement is needed, or filed into my systems, ready to compound.

I used to be the hallway between my own agents, but a new world is coming into view where the queue is where the agents co-ordinate.

My personal setup is at the do-it-yourself end of the spectrum, and it already caters for the possibility that when my colleagues have their own agents ready and able to accept work, they can monitor the same collaborate around a shared queue. But the landscape of commercially available offerings is maturing quickly. In June Anthropic launched Claude Tag, which takes the opposite approach to a queue: rather than moving work out of the conversation, it brings the agent in. Anyone in a Slack channel can tag @Claude and delegate a task, and there's one shared Claude per channel that everyone can see working. Anthropic's own word for it is "multiplayer". They claim 65 per cent of their product team's code is now created by an internal version of it. Last week Block released Buzz, an open-source workspace where humans and agents collaborate as peers, each agent carrying its own portable, cryptographically verified identity in an attempt to build the open plumbing for human–agent work rather than another walled garden.

When the network reaches the org chart

If all this stayed at the level of tasks and tickets, it would be a productivity story. In March, Jack Dorsey and Roelof Botha published an essay called "From Hierarchy to Intelligence" that makes clear it won't stay there. And before dismissing it as Silicon Valley theatre, it's worth remembering who's writing. Dorsey co-founded Twitter, which despite all its present-day shortcomings, genuinely changed how the world communicates in real time.

Their argument runs deeper than most AI commentary. Hierarchy, they say, is not a fact of nature; it's a technology - an information-routing technology, built around a single human constraint. One person can effectively manage somewhere between three and eight others, so scale forces layers, and layers slow information down. The Roman army ran on it: eight soldiers to a tent, eighty to a century, five thousand to a legion. And for two thousand years since, every organisational innovation has been a workaround for the same tradeoff. What managers in the middle actually do, in Dorsey's words, is route information - aggregate context from below, relay decisions from above, keep the parts aligned.

Dorsey's claim is that AI can now do the routing. When work is remote-first and machine-readable – by virtue of all decisions, discussions, code, plans etc being recorded - an AI can maintain a continuously updated model of what's happening across an entire company: what's being built, what's blocked, what's working. The context a manager used to carry up and down the chain, the system carries instead. Block is restructuring itself around this idea, normalising down to three roles: deep specialists, people with direct responsibility for cross-cutting problems, and player-coaches who build alongside the people they develop. No permanent layer of middle management.

Fair warning about the source: this is Block's chairman writing with his own investor, and even they concede parts of it will likely break before they work. It's conviction, not evidence, at this stage.

So what happens to us?

I work with a lot of people who live in the middle of org charts, and I want to be straight about what this argument means and doesn't mean, because the headline version "middle management goes extinct" is both scarier and less interesting than what Dorsey actually wrote.

The essay doesn't abolish the work that managers in the middle do. It splits that work in two, and only automates one half. The half the network of agents takes is the routing: the status meetings, the alignment sessions, the relaying of context between layers, the being-the-person-who-knows-what's-happening. If we're honest, that half was never the part of the job anyone loved. It was the part that expanded to fill the week and crowded out the rest.

The other half is the part the network of agents cannot do, and it's striking that even Block, in its most radical vision, couldn't design the humans out of it. Someone has to hold what the organisation is actually for, and keep people clear on it - because a world model can tell you what's happening, but it cannot tell you what matters. In a system that moves faster than any human can review, the quality of intent going in becomes the whole game, and stewarding that intent is a human role that gets more important, not less. And someone has to grow people. Dorsey's player-coach is a manager relieved of the routing work but still responsible for craft and for the humans around them - with the discipline that this demands: developing people without becoming the bottleneck the network was built to remove.

The feeling I am left with from all this is that what is old will somehow become new again – i.e. holding a vision, growing people and aligning them to bring their uniquely human skills to the fore is going to be what is left for us to do once the connected agents take care of the rest. Which is what the essence of building teams and businesses and social movements has always been about.

Where does that leave you?

The sneakernet era didn't end because we got better at walking floppy disks down the hall. It ended because the network arrived, and it arrived faster than almost anyone had planned for. Within a decade, the idea of physically carrying information between computers went from being simply how things worked to being a story you told younger colleagues who didn't quite believe you.

I think about that when I look at the gap between how remarkably capable AI systems already are and how clumsily they still share. The network for this era is being laid right now - by Anthropic, by Block, and by tinkerers like me with a free Linear account and a spare weekend. Given the breakneck speed at which the whole AI space is evolving, it will be months, not years, until the Claude or ChatGPT that currently feels like your own personal journal starts to feel something more like email.