As an example: in procurement the unit of work is not “one agent completing one task for Bob”. It is thousands of small, repeatable decisions and actions executed with audit trails. Agents, and Agent Swarms, can’t keep up with that. So what does?

Why Current Agent Metaphors Fail

Most implementations of Agents today rely on hierarchies (manager agents overseeing worker agents) or swarms pursuing one goal at a time. These metaphors break when we scale up. Error rates compound: each supervisory layer adds mistakes and false positives. Or worse, they create a massive token overhead for supervision and correction. Creating a negative ROI. You’d be better off hiring more interns.

The fundamental problem is treating agents as the core primitive is the wrong way to scale.

We do not need to “organize twenty agents”. Maybe a bakery can make a lot of profit with that. But we need to run 300,000+ reliable transactions. Stochastic parrots won’t scale for that.

The Right Primitive: Deterministic Transactions

The fundamental primitive must be audited, deterministic transactions. Reliable, compiled steps with clear inputs/outputs. We call this Compiled AI.

Agents need to switch roles as the scale increases. They translate user intent, route exceptions, and serve as human-interface glue.

What we want isn’t a roman legion of LLMs (a popular metaphor for organizing agents into groups), what we want is something like a steering wheel. A basic human interaction, that translates into a lot of reliable machine actions.

The Computer Science Solution: Orchestration as the Operating System

Computer science solved this decades ago through orchestration. The best metaphor for this is an operating system.

In a traditional Operating System like windows or iOS the following happens:

• Users interact with apps.

• Apps request actions from the OS.

• The OS schedules, validates, and executes millions of low-level operations reliably.

We are building the same thing for enterprise work: an Enterprise Operating System, we call it a Process Execution Engine.

• Top layer: Human users give high-level intent.

• Middle layer: The orchestrator schedules, validates, secures,and executes millions of low-level operations reliably.

• Bottom layer: Deterministic, auditable transactions (Compiled AI).

Unlike agent hierarchies, this contains errors in typed, inspectable, reversible transactions instead of letting stochastic agents supervise each other, and praying it will work next time.

How It Works

1. User expresses high-level intent.

2. Orchestrator maps it to sequences of deterministic transactions (which we call Plans).

3. System executes at massive scale with built-in controls.

4. Only true exceptions and decisions surface to the human.

If you’re “AI-pilled” you might start asking why there isn’t another AI above this? But there is no need for an “orchestrator of orchestrators”, at the highest level the problem becomes human alignment on priorities and tradeoffs.

That’s your company strategy, and needs to remain in the hands of humans.

Why Orchestration is a better metaphor

Thinking about AI this way scales cleanly to enterprise volumes. Want to to 100 tasks? Fine. How about 30 000 000? Yeah, no problem.

This maintains the enterprise governance we’ve spent decades creating in IT. It shares the same observability, and audibility you’d expect from any modern software.

It keep humans focused on the strategy and judgment, where it matters. And not on maintaining the stability of AI agents.

The future of work

We think orchestration can transform work. Every action a human wants to do becomes a reusable transaction. Tasks and processes are discovered and mapped as a side effect to people doing their normal work. And no one is left behind in “searching for documents hell” again.

The future is not autonomous agent armies. It is an Enterprise Operating System, a Process Execution Engine, that turns reliable transactions into a orchestrator of work at scale. We already know how to build this in computer science. Now we are applying it to organizations.

That is how we think AI will reshape enterprise work.

Subscribe now

I think what you want is your agent to go away and do really good work for you, and then come back with its work product to let you review, steer it, decide everything. But you can’t be overwhelmed, because it’s gonna do so many more things than you have time to look at, and if you’re always looking at it, you know, it can be slower than if you just did it yourself. And so it actually strikes me as an interface problem.

— on A Cheeky Pint Dario had this to say about the future of AI

The software we currently use is based on the idea of direct manipulation. It’s easy to describe as “you move a file into a folder by dragging the file into the folder”. Software wasn’t always like this. In fact a lot of developers today still use command line interfaces where you input commands and get an output. CLIs are more like indirect manipulation. But what does this have to do with AI? This interaction model needs to change, because AI breaks the assumptions direct manipulation rely on.

AI models are unpredictable, they are non-deterministic, and much slower than traditional code, which makes it impossible to use them for direct manipulation. We can use them for indirect manipulation (photo from photoshop), for example using Photoshops new magic eraser tool you can select the thing you want to remove from a photo and ask AI to do it. But you can’t just drag the thing out of the image. AI is fundamentally unpredictable, which makes it unsuitable for direct manipulation.

If direct manipulation doesn’t work, how should we work with AI? Here’s another clue from Dario’s quote, he talks about an example where you have several AI Agents doing work for you. And your job is to guide them, to help them understand when they’re doing something right and when they’re getting it wrong.

The more Agents you have working for you, the better! But the more Agents you have doing work for you, the easier it is to get overwhelmed.

At INXM we’re currently experimenting with ways of solving this, and no doubt we will find even better ways of working with AI over time. But we’ve already made meaningful improvements by building our platform to be asynchronous first. You don’t sit there watching AI writing a wall of text. When it needs your input, it will come to you.

You ask INXM to do something and the platform creates a plan to reach that goal. It tells you about it, and you give feedback. This process can happen once or ten times until you’re ready to let the plan be executed.

While INXM is doing work for you the same feedback loop continues happening, if something doesn’t go according to plan the platform will ask you for feedback or next steps before it continues working.

When INXM is done working for you, it again comes back to ask for your input.

The key here is that we treat all work as fundamentally asynchronous. The platform does work for you, and then asks for input. Whether that’s while it’s figuring out a plan to do something complicated, or checking your CRM notifications, or collecting feedback from the department heads. The specific task doesn’t matter. You simply ask it to do work, and you can trust that it will come back to you with whatever the result is. This workflow is a shift from direct manipulation to guided orchestration, the interaction model AI actually requires.

This doesn’t stop you from checking on long running work however, everything is transparent in INXM, there’s no magic black box where some agent can run amok. If you wonder why we’re still waiting on the sales report, you can just open the task and see what we’re waiting for. Steve didn’t answer? Ask INXM to reach out again.

Summary

We were happy to learn Anthropic aligns with us on what the future of work looks like. But we suspect most of the world is starting to have the same realisations we’ve been working on for the last year.

Come with us as we reinvent how we do work.

I still see people fall for this, especially each time new models are released.

AI is such an impressive technology it’s easy to start believing that it can think. It can hold a conversation better than I can, it knows more than my college professor. Maybe it actually is intelligent?

But behind the scenes, it’s just technology. Ordinary technology, just very impressive.

“Any sufficiently advanced technology is indistinguishable from magic” — Arthur C. Clarke

There is no meaning-making happening in an LLM. It’s a static mathematical model (static as in it does not change) that predicts what you want to hear, based only on language.

By choosing a specific word or inflection you can make the model behave in any way you want. That’s the magic part, not in a defined number of ways, but in Any way.

You can make it become a raving lunatic, or give you completely incorrect math, or even dangerous medical recommendations.

The companies training these amazing models claim they are in a race to create an Artificial General Intelligence. A super intelligent model that can learn new things. A super human intelligence. Or at least that’s the story they’re selling to us. Because there is zero evidence of this.

How smart is a model that cannot change? No LLM can. This isn’t a small thing, we need totally different models for them to be able to learn. How smart is a model that can be persuaded to agree on bad math?

It’s not smart. Not at all. But it is phenomenal at language. It’s so good, in fact, that we start believing it knows all sorts of things. So good we start believing its magic.

To make useful tools we need to make them understandable first. A calculator or a pair of scissors has a specific function. We understand what function that is, so we can use them effectively, even if we couldn’t explain exactly how a calculator calculates.

At INXM, one of our goals is to make our Orchestrator understandable. While the user can chat with the platform, we’re not trying to pretend it’s magic. It isn’t, it’s just an interface that fully understands your language and what you mean.

INXM Orchestrator can perform the same tasks you can, using the same software and tools. The difference is it does it 24/7, it doesn’t forget, and if something unforeseen happens, it asks you what to do.

It’s not magic. It’s just smart code. But when it asks you what to do next, it feels almost magical.

It can seem like AI is flooding us with slop. But this isn’t a new problem, just an old problem at a higher velocity.

Text as a medium

Text is a magical medium. It lets us send information forwards in time. It can communicate complicated ideas. And it’s very information dense. The right words can communicate an entire world in a paragraph.

Compared to spoke language, reading and writing has a much higher rate of signal to noise.

But text is also demanding. Writing takes a lot of effort. There is even some evidence that rendering fuzzy thoughts into the right words is the active form of thinking. This makes text less great for moments when you need a lot of actions.

Thinking: super fast but fuzzy
Speaking: fast but vague
Writing: slow but specific

This tells us something about how to design better tools for work.

Thought experiment: designing better tools

Imagine if your car was self driving but you would have to manually text it where to go: “Right, Left, Left, Right”. That keyboard would be pounded to bits.

What about voice? Less information dense and takes less effort. Better in the car too. But you’d still get very tired of saying “Right, Left” quite soon.

The better way is to shape the context of information around what the user is trying to do. In a car they always have a destination, so a natural language UI can simply parse that, and then use a routing service to understand where to go. Close to zero effort for the user, they can verify on a map that the car got it right or wrong, no need to read a long text.

“Context is the interface The best agents don’t stay in isolated chat dialogues. They move into surfaces we already trust. “I feel like a browser … because you can do an infinite number of things on a browser and we’re all using browsers.”
— Liam Matteson, Browserbase

This is how we thinking about the future of work.

What is the context our user is in? She’s between meetings, in her email, chatting in teams, or in a face to face conversation. She’s trying to reference data, or complete some process, in her companies software suite.

Anything from “create a new deal flow pipeline in our CRM”, and “look up the material costs of regional projects”, to “Did procurement respond about this request from last week I don’t remember?”

Each request can be made simply through natural language, or through a specific User Interface depending on the users context.

Coming back to our car analogy: INXM Orchestrator understands enough about the landscape she’s in to realise which systems are needed to perform this journey.

Orchestrator then presents that specific UI, in this case a map, to the user so she know where she’s going. And can even add extra stops along the way if they need.

When she accepts, Orchestrator fires off this journey plan as a series of tasks in the systems and programs she normally uses. The map shows her the current progress. And if anything changes Orchestrator will ping her to let her know.

The future of Ai, and of software, is not about reading and writing more text. It was always going to be about showing you information more intelligently. Making it easier for the human user to take the right actions at the right time.

For INXM this means we spend an absurd amount of time testing what information is right for which actions. What is the “map” for something as complicated as “material costs of regional projects”?

This future of work wont be perfected any time soon, but we’re excited about the steps we have already taken. And we cant wait to see you put more work on cruise control.

Before Google the amount of energy it took to look something up made it less likely that you would try. Then along came Google, and suddenly everyone googling the most random things. Why not? It just takes a few seconds. This is what I think about when the headlines try to sell that AI is going to take our jobs.

AI will make a lot of the work we’re doing today laughably easy to do. Things we used to charge for, things that took us days, will now take cents and seconds. But do you think that means we’ll run out of things to do? Google made a lot of library indexes obsolete, and created a flood of new opportunities. AI will do the same.

Think about the number of clicks you make on your computer in a day. It’s in the thousands. Each one didn’t take long, but I’m willing to bet you still sigh from time to time before you open your last notification. This friction, the same friction we felt to look something up, can now go away with AI.

There’s simply no need to deal with many of the incoming messages anymore. AI will do it. You can focus on the actions the message thread is trying to coordinate.

The next time you set up a process for procurement, there’s no need to train people, or document it. Just ask AI for a draft and have it ask each stakeholder for feedback to incorporate. AI will create the process. You can just procure.

Productivity increases have unintuitive outcomes. We think they mean that demand will dry up: How many shoes can you possible wear? But that’s not what happens. As supply increases we can be much more creative and do more interesting things: Sneakers are not collectibles as much as foot protection.

As employees become more productive we might hire less, for a while. But when we’re more productive, there’s more money to invest. And we need a lot more people for all these new ideas.

We’re never going to run out of problems to solve, or things we want. So the amount of work that to do will always increase, never decrease.

I can’t wait to have an AI army working for me. I’ll need to hire a lot of people to try out all of these ideas.

Subscribe now

Share