For decades, every piece of software was an island.
To use one, you rowed across and paid the entrance fee, and then you were admitted: you got a login. Then you learned the local language: you needed to understand where the menus lived, what each button did, how to make exports or conduct searches within the software.
- Each time you got a new CRM, you needed to learn a new set of rules — new dashboards, new filters, new names for the same objects and relations.
- Each time accountancy changed their invoicing tool, you had to understand what the folder system was like, where the archive was and how to retrieve data.
- Each time the fine folks at procurement set up a new piece of software for their processes, you needed to adjust: how to enter your project, how to keep track, where to sign off for what, etc.
And so it went, island after island, each with its own jargon, its own dialect. Island after island. Software after software.
Nobody complained about this.
There simply was no other way to reach what the software could do, so you made the crossing and played by its rules.
That era is ending.
The bridges are arriving
The change rests on two things:
- The first is that you have now gotten used to working with at least one, if not multiple Large Language Models: ChatGPT, Claude, Copilot, Gemini or another one. It's convenient, it's very capable, it's always on, and that's why it's quietly becoming the place where you work. You open it in the morning, you think in it, you hand it your tasks. It is no longer a clever toy beside your real tools. For more and more people it is the mainland, the one place they return to all day.
- The second thing is something called the Model Context Protocol, or MCP. An MCP is a bridge. It connects a single island — one piece of software — to the mainland: to the Large Language Model. And it can build bridges from your LLM to any software you care to work with.
Any bridge that is built stays forever. The LLM can from then on use those bridges to reach software, collect what it needs, and you never have to learn the software's jargon or idiosyncrasies to get the benefit of the software.
Put one and two together and the old way of working changes. You used to spend your days skipping from software to software, getting information from each of them, and trying to reach your objectives with all of that info.
Now, you can stay inside the LLM you already know, and tell it what your objectives are. That LLM will then decide what software it needs to help you, reach across to the right software, and bring the info you require back to you. Even without you requesting that info explicitly.
What the software is capable of has not changed. How you unlock that potential has: this has dramatically improved. The ease with which you operate has increased massively.
What happens once the bridges open
Two things follow, and together they are larger than either alone.
- The interface disappears into the model. You don't have to learn how to use each software. You say, in plain words, what you want, and the LLM goes and does it. The hours once spent finding where things live and learning about how to find them simply vanish.
- Software begins to combine on its own. This is the part that makes it all truly mesmerizing. The LLM can walk across several bridges at once — one capability from here, another from there, a third from somewhere else — and assemble all of the info into a result none of those products could have produced alone.
What it means to build software now
It would be easy to read this as bad news for software. If the LLM does everything, who still needs the tool underneath? But consider what the large language model actually needs to work well.
An LLM is at its best when it reasons over information that is already clean, structured and ready to use.
Ask it to wade through raw data itself — to gather, sort and make sense of the mess — and it slows down and stumbles.
Someone has to hand it the finished, valuable material.
That is what good software does, and has always done: it turns raw data into something dependable and ready. The LLM feeds on that. Without it, it starves: it starts to hallucinate, just like someone who is starving does.
Which is why software does not become redundant here. It becomes the thing waiting on the far side of the bridge with the value already prepared.
This is the end of the era of islands, not the end of the islands themselves. They stop being places you had to cross the water to reach, and become the sources the mainland cannot do without. The model can cross any bridge it likes — it still arrives empty-handed unless someone, on some island, has the goods ready.
Software will not be replaced by AI. It will be engulfed, embraced, to all appearances gone.
And yet, it will matter more than it ever did.
