Over the years, my home has become something of an Island of Misfit Kit because it’s filled with ancient computers, all of them well and truly out of support.

Most still work but it was always inadvisable to use them, and more so since Mythos came along.

So I mostly left them alone until a few months ago when I decided to see if an agent could help to revive a 15 year old device with a tiny CPU, an ancient version of Arch Linux , and a browser so old it lacks the necessary encryption to connect to the modern web.

I let the agent SSH into the device, gave it full control (yes, yes, but I know what I'm doing, I promise) and then watched as over three hours the agent repeatedly failed to get a toolchain that would compile the latest cURL with modern cryptography.

The agent finally did the job, albeit at glacial speed because this ancient CPU lacked even rudimentary floating point capabilities.

I eventually asked the agent if someone else had solved this problem? It quickly found a fast math library, and got cURL purring. A few minutes later I had my agent on that device, chatting to me via Telegram.

If an agent could do that, what could it do for the rest of my retro collection?

And just like that, the agents invaded. One minute I was wondering "can I get some help with...?" and the next minute they were everywhere, on everything, like the Weeping Angels of computing.

A resuscitated 12 year-old iMac Pro, long past its use-by date, now happily running the latest Ubuntu with all of the correct firmware tweaks, installed and maintained by an agent. Ditto the 10 year-old virtual reality PC, and a Surface Go stuck at Windows 10 and underpowered at birth. All working beautifully now, because as soon as the OS boots and the network link comes up, I install an agent.

I could do these configurations by hand – but who has time to read through all the Reddit posts and tech notes and GitHub repos? Isn't that sort of thing best left to machines, precisely because it leaves me free to use these machines in all sorts of interesting ways?

Now that it's trivially easy to feed the parameters for any brainwave into an agent, I've gone full mad scientist, burning through ideas precisely because the cost of testing them has almost completely disappeared. I'm learning at a rate I'd never been capable of before, feeding a rough idea in, then leaving it to the agent to loop through multiple paths for however long it takes to get a 'good enough” result. I propose, agents dispose – then hand over the results.

That's not just the future of retro-tech revivals; it's coming for pretty much all the hard sciences. Four of the key personnel at Google DeepMind quit last week to launch Discovery Loop, applying these same practices to biology, physics, materials science, and much more besides. Scientists propose; the agents run the tests.

Discovery Loop is the shape of what's coming – not just in science, but in business, in governance, and in pretty much every other domain where you can measure your results. The trick there isn't the automation; it's understanding which measurements matter, which can be gamed, and which should be ignored.

That's not something you can learn from an agent, because all of that qualitative judgement resides in the human domain. Agents have given us speed and scope, and highlighted the real bottleneck. It isn't compute or RAM or even ideas worth testing; it's the judgement that lets us decide what's best. ®