Gareth Ransome

Work, rugby, solve a problem, have a beer, go on the boat

[Gareth Ransome @GarethRansome]

Most admin systems have a problem nobody talks about: the rules of the business are baked into the software itself. The process lives in the code. So when the business changes — new step, new rule, new type of customer — you've got a software project on your hands. Developers, testing, weeks or months, and the risk of breaking something else. It's why most companies end up bending their business to fit their software instead of the other way round.

I built mine the other way. Twenty years of it.

The idea is simple to say: everything that can change is data, not code. The screens, the steps, the decisions, and what happens after each decision — all of it lives in tables, like recipe cards in a box. The software is just an engine that reads the cards. When the business changes, you don't rewrite the engine. You edit the cards. That's why any change — literally any — lands in days, not months. Nothing gets rebuilt. Rows get edited.

In practice it's three layers, and you don't need to be technical to follow them.

Task sheets. Every job in the business is a numbered list of steps that anyone can follow. No tribal knowledge, nothing living in someone's head. The sheet carries the process so the person doesn't have to.

Decision gates. At certain points in a sheet, a person makes a judgment call — and they make it by picking an outcome from a fixed list. "This is a new site." "This needs investigating." "Send it to accounts." That's the whole human job: pick the right branch.

Outcome tables. Here's the clever bit. The moment an outcome is picked, the system fires everything downstream automatically — updates the records, sends the template email, logs the history, creates the next task for the next person. Every one of those consequences is just a row in a table. Want the outcome to do one more thing? Add a row.

So over time the human's job shrinks to pure judgment at the gates, and everything mechanical runs itself from configuration.

Now — why does this matter for AI agents? Because an agent doesn't need to be clever to work in this system. It sits in exactly the same seat as a person: reads the task sheet, follows the steps, and picks an outcome from the defined list. It can't go off-piste, because the outcomes are the only levers that exist — and the consequences of each outcome are fixed in the table, not improvised by the AI. The safety boundary isn't a policy document. It's the architecture.

And here's the part that still makes me smile: I spent twenty years writing step-by-step instructions so any member of staff could do any job. Turns out a clear instruction sheet for a human is a prompt for an agent. I was writing agent prompts for two decades without knowing it.

Developing it further for agents is just more of the same method. Start the agent on one small task and watch its outcome picks. Add an "escalate to a human" outcome for anything it's unsure about — escalation is just another branch, no special machinery. Automate more rows so the sheets get shorter. And when you want the agent doing something brand new, you don't write code — you write a new task sheet and a few outcome rows. Minutes.

The last thing I'll say: there's nothing about signs in any of this. Every admin job in every business has the same shape — information arrives, someone works through steps, makes decisions at branch points, and things happen as a result. Change the cards in the box and it's an insurance system, a legal system, a logistics system. The engine doesn't care.

That's what most admin systems don't do. And it's why the agents, when they arrive, will find the seat already built for them.

[Gareth Ransome @GarethRansome]

You can’t force people to employ someone if there is no money… and how do you compete with a startup who replaces no one, just competes with no staffing cost…. But I agree it’s not going to be a comfortable change, but change it must and it’s going to hurt some and benefit others… such is the way of things… I’m doing my best to ensure the people I employ have the best chance. If a government stops me making changes I don’t have a choice about it will be worse….

[Gareth Ransome @GarethRansome]

@MathieMyers @alexandr_wang Then if they do they will fail…. Because expectations will rise and they will not be able to follow… I implement AI in work processes and it still needs people and will do for sometime…

[Gareth Ransome @GarethRansome]

AI will change not replace…. I employ people and am not thinking of replacing I am thinking of how we do more… because we will not be able to charge as much when it is easier to do so we will need to do more… it seems counter intuitive but I can see how it will affect my businesses, governments needs to keep there noses out, they interfere and it will be a nightmare, no one has a clue what’s going to happen, least of all clueless bureaucrats

[Gareth Ransome @GarethRansome]

@paradite_ Or you do what had been done for decades use processes to do it for you…..maybe? Why reinvent the wheel? This is why company’s ISO

[Gareth Ransome @GarethRansome]

Hi Michael, I've sent you the detail by DM, but the short version: I run a small signage firm, and I'm not going to build an agent harness because I think I already have one. Every job we do is a task sheet, and each sheet already has a table saying what a person may read and a form saying what they may write. My view is that agents go wrong the same ways new staff do, so I'm giving the agent the same table and form for its task, as data, in its own name, and it's gone when the task closes. Nothing new, just the constraints we've used on people for years. Would be interested whether your API would help in implement that? We have just been building the MCP yesterday and today…….. so the timing could not be better…..

[Gareth Ransome @GarethRansome]

@mark_k Could not agree more 👍

[Gareth Ransome @GarethRansome]

Your thoughts on this idea 😊but maybe it would only be if interest to me…

I think we already have AGI. The argument about whether we do is stuck because it’s comparing the wrong things: a person to a model. Neither of those is the unit that matters. Both are parts. What’s intelligent, in both cases, is the system.
Take a human. On its own it can’t feed itself, can’t reproduce, can’t learn language, can’t build a fab. It survives because it sits inside a network: bodies + microbiomes + kinship + agriculture + language + institutions. Drop one adult in the wild with none of that and you get a corpse in a fortnight. What we call human intelligence is that whole assemblage, with the brain as one inner loop inside it.
Now take AI. A model on its own does nothing either. It sits inside models + agents + researchers + data pipelines + clusters + capital + chip and energy supply chains. Same shape. An outer loop (training runs, experiments, paper-and-code, capital allocation) and inner loops (individual models). It pulls electricity, water, minerals and human attention from its environment the way we pull sunlight through plants.
Once you compare system to system, the trait-by-trait scoring falls apart. “Can it reproduce?” A human doesn’t emit another adult; the population plus care plus culture does. The AI stack doesn’t emit another industry; labs plus models plus markets do. In both the reproducing unit is the network. Ours is good at copying bodies and slow at copying infrastructure. Theirs is the reverse: it copies its digital layer almost for free and its physical layer barely at all.
So the useful question isn’t “is it alive” or “is it AGI”. It’s: how closed is the loop, how self-directing, and how substitutable for its human parts is it becoming? On that measure the stack is a young, fast, incomplete organism. Strong at elaborating itself digitally. Still dependent on human institutions and industrial metabolism for embodiment, energy and purpose.
And “dependent” won’t stay one-directional. Mitochondria were free-living bacteria once. Endosymbiosis doesn’t end with the guest leaving; it ends with the guest becoming part of what the cell is. Whether AI is being absorbed into the human network or humans into the AI one depends entirely on where you draw the membrane, and that’s a choice, not a measurement.
One more thing that gets missed. People think the gap is hardware: robots, fabs, power. I think the hardware is the easy half. The real closure gap is that the human network keeps its genome in tacit knowledge. Every process that lives in someone’s head is a bit of metabolism the stack can’t copy. Every one that gets written down as structured rows crosses into the digital layer, where the stack already wins. That boundary is moving every day, and it’s moving because we’re moving it.
The “is it AGI” test people want, a single agent you can drop into an open world that keeps going, fails for humans too. What it actually measures is how much of the network an individual can carry compressed. A human carries twenty years of it. A model carries a checkpoint. Different ratio, same test.

[Gareth Ransome @GarethRansome]

@AlexPeire @EMostaque I’ll try that too…. The first suggestion seems promising too… need to try it a bit more extensively but it definitely seems better, the question is can Claude improve the output over several turns I will be finding out… thanks for taking the time to help, really appreciated 😊

[Gareth Ransome @GarethRansome]

@AlexPeire @EMostaque Hhhhmmmm right ….. I use both, but don’t put one in the role of checking the other… that’s a great idea….. yeah thanks for that I’ll give that a try 😁

[Gareth Ransome @GarethRansome]

I never hard code anything if I can avoid it… we soft code where engines run database entries to output anything for our system / business we are getting rid of the stuff where it is hard coded as it kills flexibility for the actual operation of the business, which is my actual goal as the business owner…. But I get what you are saying I have formalised what it can do to some degree and get some benefit…. But it seems like it ought to be able to do more….

[Gareth Ransome @GarethRansome]

Yeah , you are describing what I have to do at the moment, but I can do it faster myself as I can do all at once, then the model ends up being a fancy spell check that points out issues that are not issues just to feel like it’s being useful and I have to ignore those… I have eventually got enough detail that is can help set up most sheets so they are roughly correct and finish and convert them to a databased version but it’s the one job where it is not great and it’s the job where until it’s able to do these skills, it will never be able to plan and act at a certain level Ina company…..

[Gareth Ransome @GarethRansome]

@K8050906068853 @EMostaque Released models.

[Gareth Ransome @GarethRansome]

@EMostaque I am four weeks in improving a workbench it’s not bad, not terrible, but it just misses the point and I have to do the last 30% all the time.. and it’s not that it doesn’t know it knows just doesn’t apply it at the right time or it applys just bits never it all

[Gareth Ransome @GarethRansome]

Then why does it really struggle with business processes??? any advice on how to get it to be able to perform better? Both latest models are the best I have seen and yet still quite poor even with huge amounts of context, it confirms it has the context, just only applies certain bits of it, it’s like it can’t hold all of it at once, reminds me of coding 12 months ago, so I assume fixable….??? Any help appreciated 😔

[Gareth Ransome @GarethRansome]

I think we already have AGI. The argument about whether we do is stuck because it’s comparing the wrong things: a person to a model. Neither of those is the unit that matters. Both are parts. What’s intelligent, in both cases, is the system.
Take a human. On its own it can’t feed itself, can’t reproduce, can’t learn language, can’t build a fab. It survives because it sits inside a network: bodies + microbiomes + kinship + agriculture + language + institutions. Drop one adult in the wild with none of that and you get a corpse in a fortnight. What we call human intelligence is that whole assemblage, with the brain as one inner loop inside it.
Now take AI. A model on its own does nothing either. It sits inside models + agents + researchers + data pipelines + clusters + capital + chip and energy supply chains. Same shape. An outer loop (training runs, experiments, paper-and-code, capital allocation) and inner loops (individual models). It pulls electricity, water, minerals and human attention from its environment the way we pull sunlight through plants.
Once you compare system to system, the trait-by-trait scoring falls apart. “Can it reproduce?” A human doesn’t emit another adult; the population plus care plus culture does. The AI stack doesn’t emit another industry; labs plus models plus markets do. In both the reproducing unit is the network. Ours is good at copying bodies and slow at copying infrastructure. Theirs is the reverse: it copies its digital layer almost for free and its physical layer barely at all.
So the useful question isn’t “is it alive” or “is it AGI”. It’s: how closed is the loop, how self-directing, and how substitutable for its human parts is it becoming? On that measure the stack is a young, fast, incomplete organism. Strong at elaborating itself digitally. Still dependent on human institutions and industrial metabolism for embodiment, energy and purpose.
And “dependent” won’t stay one-directional. Mitochondria were free-living bacteria once. Endosymbiosis doesn’t end with the guest leaving; it ends with the guest becoming part of what the cell is. Whether AI is being absorbed into the human network or humans into the AI one depends entirely on where you draw the membrane, and that’s a choice, not a measurement.
One more thing that gets missed. People think the gap is hardware: robots, fabs, power. I think the hardware is the easy half. The real closure gap is that the human network keeps its genome in tacit knowledge. Every process that lives in someone’s head is a bit of metabolism the stack can’t copy. Every one that gets written down as structured rows crosses into the digital layer, where the stack already wins. That boundary is moving every day, and it’s moving because we’re moving it.
The “is it AGI” test people want, a single agent you can drop into an open world that keeps going, fails for humans too. What it actually measures is how much of the network an individual can carry compressed. A human carries twenty years of it. A model carries a checkpoint. Different ratio, same test.

[Gareth Ransome @GarethRansome]

I think we already have AGI. The argument about whether we do is stuck because it's comparing the wrong things: a person to a model. Neither of those is the unit that matters. Both are parts. What's intelligent, in both cases, is the system.

Take a human. On its own it can't feed itself, can't reproduce, can't learn language, can't build a fab. It survives because it sits inside a network: bodies + microbiomes + kinship + agriculture + language + institutions. Drop one adult in the wild with none of that and you get a corpse in a fortnight. What we call human intelligence is that whole assemblage, with the brain as one inner loop inside it.

Now take AI. A model on its own does nothing either. It sits inside models + agents + researchers + data pipelines + clusters + capital + chip and energy supply chains. Same shape. An outer loop (training runs, experiments, paper-and-code, capital allocation) and inner loops (individual models). It pulls electricity, water, minerals and human attention from its environment the way we pull sunlight through plants.

Once you compare system to system, the trait-by-trait scoring falls apart. "Can it reproduce?" A human doesn't emit another adult; the population plus care plus culture does. The AI stack doesn't emit another industry; labs plus models plus markets do. In both the reproducing unit is the network. Ours is good at copying bodies and slow at copying infrastructure. Theirs is the reverse: it copies its digital layer almost for free and its physical layer barely at all.

So the useful question isn't "is it alive" or "is it AGI". It's: how closed is the loop, how self-directing, and how substitutable for its human parts is it becoming? On that measure the stack is a young, fast, incomplete organism. Strong at elaborating itself digitally. Still dependent on human institutions and industrial metabolism for embodiment, energy and purpose.

And "dependent" won't stay one-directional. Mitochondria were free-living bacteria once. Endosymbiosis doesn't end with the guest leaving; it ends with the guest becoming part of what the cell is. Whether AI is being absorbed into the human network or humans into the AI one depends entirely on where you draw the membrane, and that's a choice, not a measurement.

One more thing that gets missed. People think the gap is hardware: robots, fabs, power. I think the hardware is the easy half. The real closure gap is that the human network keeps its genome in tacit knowledge. Every process that lives in someone's head is a bit of metabolism the stack can't copy. Every one that gets written down as structured rows crosses into the digital layer, where the stack already wins. That boundary is moving every day, and it's moving because we're moving it.

The "is it AGI" test people want, a single agent you can drop into an open world that keeps going, fails for humans too. What it actually measures is how much of the network an individual can carry compressed. A human carries twenty years of it. A model carries a checkpoint. Different ratio, same test.

[Gareth Ransome @GarethRansome]

I’ve got one it really struggle with any advice on how to get it to be able to perform better? Its business process and it’s the best I have seen and yet still quite poor even with huge amounts of context, it has the context, just only applies certain bits of it, it’s like it can’t hold all of it at once, reminds me of coding 12 months ago, so I assume fixable….

[Gareth Ransome @GarethRansome]

I find it amazing and still in the area I work, in the hardest task I do, creating processes to run my company, it is poor still….

It’s better than anything that’s come before but even with me lining the context up in a lot of detail and giving it a template to copy and giving it all the rules, it still gets overwhelmed and fails to produce a good output….

It had the context and just didn’t access or apply it….. it has said yes I can see I had that and this is this guiding principle and I should have applied it etc….

So not AGI….. amazing when I use it to code, magical almost……but when it comes to business processes it fails and honestly looks a little flaky, like coding from 12/18 months ago…..

[Gareth Ransome @GarethRansome]

I moved over to GPT for a project yesterday and changed my subscription.

Until this point Fable was my only option…. GPT 6 is better.

Business process design and recording is the project and until now all models have struggled even Fable… 5.1 is better but GPT 6 can actually be useful without me holding its hand… that’s a big change in a week