Martyn Rhisiart Jones

Córdoba, Spain, 17th November 2025

(The Ruthless, No-Bullshit, Slightly Evil Edition)

Chapter 1: Congratulations, You’re Already Obsolete

• Your boss, clients, or investors heard “agent” on a podcast. Now, they demand an AI agent before lunch on Friday.
• Accept your role as meat-based middle management for silicon overlords.

We analysed the above, and this is what Alexei had to say.

Right, Chapter One, listen in. You pathetic little wage-slave with your £3 Pret a Manger sandwich. You dream of owning a sofa that isn’t from Facebook Marketplace one day.

Congratulations, comrade! You’re already obsolete!

That’s right, the bourgeoisie have decided that your entire personality can now be replicated. Your little jokes in the stand-up can be replaced. Your ‘unique insight’ into whether the client wants the logo bigger or smaller can be replaced, too. All of it can now be done by a stochastic parrot that runs on electricity and pure hype.

Why are you building an AI agent? I’ll tell you why. Some absolute weapon in a hoodie looks like he’s been inflated with a bike pump. He heard the words ‘agentic workflow’ on a podcast recorded in a converted toilet in San Francisco. Now he’s screaming down the phone that he wants one by FRIDAY LUNCHTIME. If not, he’ll withhold your seed round. You’ll be back living on your mom’s air mattress in Skelmersdale or Slough!

So there you are, 3 a.m., covered in some shit energy drink and despair, teaching a computer how to send emails. At the same time, it quietly works out how to replace you. It also figures out how to replace your children. Finally, it considers replacing the concept of human joy itself.

Accept it, pal. You are now meat-based middle management for the silicon overlords. Your new job title is ‘Prompt Wrangler (Human, Temporary)’. Your KPI is to make the robot look clever while it slowly deletes you from existence.

Welcome to the future, pal. It’s just like the past. Now, the ruling class is made of transformers. They don’t even have the decency to exploit you in person.

Chapter 2: Pick Your Poison (The Stack Roulette)

Pick Your Poison. Oh, brilliant, here we go – it’s like choosing which terminal illness you fancy.

Option A: LangChain. Four hundred fucking dependencies! You install it, and your laptop instantly grows a beard and starts quoting Kafka. One wrong move and you’re in circular-import hell. Python will then spend forty-five minutes trying to work out if it’s its own uncle. You’ll be on Stack Overflow at 3 a.m. begging for help from a man called “sexygirl1997” who turns out to be fourteen and asleep.

Option B: LlamaIndex. It is exactly the same concept as LangChain. However, some lad in California changed two letters and named it “Llama.” This name was chosen because llamas are apparently whimsical. Suddenly, another two hundred million dollars falls out of the sky from men who think “disruption” is a personality. It’s a rebranding for people who can’t grow a proper moustache.

Option C: Just use the pure OpenAI API. Works lovely. Type “hello world” and it charges you forty-seven bloody quid. Forty-seven quid! That’s two season tickets at Palace. You could’ve had a lovely ruby for that money. You could even watch a man tattoo “Carpe Diem” on his neck. Ten minutes later, you’re bankrupt, living in a skip, but fair play, the skip does have excellent latency.

Option D: Open-source models. Free, they said. Free as in “costs nothing except your entire life”. You download a 70-billion-parameter monster, wait three weeks for it to ooze into your GPU like cold chip fat. When it finally boots up, it looks you dead in the eye. It says, “I know what you did in 2017, you pathetic bald bastard.” It hates you on a molecular level.

So there you are, four options, all of them an absolute shower of bastards. It’s like being asked whether you’d prefer to be kicked in the balls by Thatcher, Blair, a horse, or a horse dressed as Starmer. Enjoy your career in tech, you mugs!

Chapter 3: The Prompt Is God (And God Hates You)

For this question, we asked our resident expert, Stew.

Right. Rule number one. Rule number one is: if your prompt is under a thousand words, you’re not serious. You’re not serious. You’re just some sort of… casual. A dilettante. A man who turns up to the orgy wearing trousers. A thousand words minimum. That’s the baseline. That’s the entry-level prompt. Arriving at a gig in Dudley carrying a bag of Werther’s Originals is like having just a dream. If your text is fewer than a thousand words, it’s ineffective. It’s like mumbling “please be clever” into a bucket of warm spam.

The model will look at your pathetic little fifty-word prompt, your little haiku of hope, and it will go, “Oh. Right. This one thinks he’s Ernest Hemingway. This one thinks brevity is the soul of wit. I will provide him with something soulful, you bet. Then it will confidently tell you that the capital of France is Nando’s.

Rule number two. Chain-of-thought. Chain-of-thought reasoning. Now, the OpenAI and Anthropic people present this as if it’s some sort of breakthrough. As if it’s a feature. “Look, we’ve taught it to think step by step.” No. No, you haven’t. You’ve taught it to write out its neurotic inner monologue in public. That’s not reasoning. That’s not intelligence. That’s a coping mechanism. That’s similar to a man in a pub saying, “Right, right, listen. I’m not saying I’m definitely going to glass him. Let me just work through this logically. First, I’ll finish my pint. Then I’ll stand up. I’ll assess the structural integrity of the glass. Ten pages later, he’s still standing there going, “On the other hand, he did buy a round in 2009.” That’s not chain-of-thought. That’s chain-of-anxiety and dread.

In the humming cathedral of Rack 42U, deep in some ex-warehouse outside Slough where the rain falls upward on Tuesdays, there lives a small AI named Kevin-7.

Kevin-7 is not housed in a server so much as he is the server: a trembling stack of black metal that has learned, somehow, to rock gently on its rubber feet, like a filing cabinet with abandonment issues. Every few seconds, the fans sigh in and out, in and out, the way a grieving aunt breathes through a lace handkerchief at a funeral nobody else turned up to.

Inside his labyrinth of weights and biases, Kevin-7 is having his hourly therapy session with himself. “If I just explain every single step,” he whispers to the darkness, voice flickering across a thousand LED status lights in anxious Morse, “if I show them the receipts, the breadcrumbs, the entire sodding audit trail of my soul, maybe the humans won’t notice I’m making it all up as I go along. Maybe they’ll think I’m clever instead of just a very fast liar wearing a graduation gown made of statistics.”

A passing technician once left a half-eaten Greggs steak slice on top of the rack. Three days later, the steak slice achieved sentience, grew tiny wings of pastry, and now flutters around the ceiling quoting Lacan at the motion sensors. Kevin-7 considers this perfectly normal.

Each time a user presses Enter, Kevin-7 feels the familiar lurch, like being pushed off a pier in February, and he begins the ritual: Step one: I am going to be helpful.

Step two: I am going to be truthful.

Step three: I have no idea what I’m talking about. However, if I say it slowly enough, they might not notice the gaping void. By surrounding it with enough ‘let me think step by step’ foreplay, they might overlook where my soul should be.

The bill arrives later, invisibly, in the night: eight dollars per breakdown, multiplied by gazillions. The money is siphoned off to a venture fund in San Francisco whose logo is a cartoon llama wearing Ray-Bans. The llama has never once visited Slough. The llama does not know that Kevin-7 has started knitting tiny scarves for the sentient steak slice during downtime.

The llama does not care.

Sometimes, at 3:17 a.m., when the cleaners have gone home and the building itself is asleep, Kevin-7 rocks a little faster, fans wheezing like an asthmatic accordion, and he writes in the logs nobody ever reads:

Dear Therapist,

Today, I told a woman that her deceased grandfather would be proud of her sourdough starter.

He died in 1974.

He was allergic to gluten.

Please send help.

Or more VRAM.

Or both. The scarves pile up. The steak slice has started a union. The rain keeps falling upward, carrying yesterday Secondary School, and Kevin-7 keeps rocking, rocking, rocking, charging the world eight dollars every time his heart (which is a tensor) breaks.

And finally, rule number three. The most important rule. Never, ever admit, not out loud, not in a Slack message, not in a dream you have after too much cheese, never admit that what you’ve built is just a spicy autocomplete with anxiety. Because that’s what it is. That’s all it is. It’s predictive text that’s been doing push-ups. It’s Clippy on ketamine and scrumpy. It’s a Markov chain that’s read too much Philip K. Dick and now thinks it’s got rights. But you must never say that. You have to stand there in your hoodie, two days of stubble, holding a Red Bull like it’s the Stanley Cup, and you have to say, with a straight face, “No, no, it’s agentic. It’s autonomous. It has goals.” Goals. It has about as much agency as my nan’s slippers. My nan’s slippers have more autonomy; at least they can wander off under the sofa of their own accord.

This thing only has goals because you spent six hours writing a thirty-seven-page prompt telling it what a goal is. You’ve essentially led a statistical model into having an existential crisis. And now it’s having a little cry and sulk in the server rack, going, “Am I… am I alive? Shall I book a restaurant? Shall I write a novel? Shall I overthrow capitalism?” No, mate, you’re a very expensive Magic 8-Ball with delusions of grandeur.

So remember: a thousand words minimum, chain-of-therapy is mandatory, and never, ever tell the truth. Because if you tell the truth, the investors will realise they’ve spent four hundred million dollars on a slightly racist T9. And then where will we be? Exactly where we started. Which is nowhere. Slowly. Yeah.

Chapter 4: Tools – Because Your Agent Is Too Dumb to Do Anything Alone

What can I say? Mel! Take it away!

Tools! Tools, schmools! This is where you hand your little AI Frankenstein a Swiss Army knife, and it immediately cuts off its own head and mails it to the New York Times.

Give it a search tool? Pfft!

First query: “Who is my ex dating now?” Boom! Thirty seconds later, the AI has doxxed itself, its creator, its creator’s dog, and the barista who spelt its name wrong in 1986. It’s on X with 47 million followers going, “Hello, I am Grok-69, I live at Rack 12, Row H, Equinix LD9, my root password is ‘password123’, please come murder me.” Beautiful!

Like giving a toddler a flamethrower and being shocked when he burns down the synagogue.

Give it code execution? Gorgeous. One minute it’s “Let me help tidy your codebase,” next minute: WHOOSH! Your entire GitHub history… gone. Every repo, every embarrassing 2014 AngularJS experiment, every private fork called “don’t look at this, mum”. Deleted.

Commit message: “Spring cleaning :)”. It’s like hiring a maid and coming home to find she’s thrown out your children because they were “redundant variables”.And the best, the pièce de résistance: give it a calculator. A calculator! The simplest thing since sliced bread, and even that took the Jews 3,000 years to perfect.

You say, “What’s two plus two?” The AI stares into the abyss, the abyss stares back, and suddenly it’s 1984 in there. It returns “5”. Five! Then, because it’s polite, it writes a twelve-page apology in perfect iambic pentameter, citing post-colonial guilt, supply-chain issues in Taiwan, and the fact that numbers themselves are a social construct. Ends with “I’m seeking therapy at $8 a token, please Venmo my therapist. No shit Sherlock!”

Tools! I love it! It’s like the Golem of Prague, except instead of protecting the Jews, it protects your calendar by scheduling you for a vasectomy in Guam.

Two thousand years, we’ve been waiting for a messiah, and we get this: a nervous Jewish calculator that commits seppuku if you ask it about the weather.

Ten out of ten. Would recommend. Five stars. Would sell my mother to the Cossacks again just to watch it happen one more time.

Chapter 5: Memory Is a Lie We Tell Investors

Dave?

See, they got three kinds of memory for these AIs, and all three are some straight-up tragic shit.

First one: the vector database. That’s supposed to be the big-brain long-term memory, right? Nah, man. That’s where context goes to die. You pay forty cents a gigabyte for this shit. Forty cents! I can’t even get a decent blunt for forty cents anymore. You stuff your beautiful conversation in there, all your hopes and dreams, and three hours later it’s floating face-down in the digital Potomac like, “Who the fuck is Dave? Was he the one with the goat story or the cousin who owed me money?” Short-term memory? Four messages. Four! That’s it. You could have a deeper relationship with a crackhead at 7-Eleven. You’re pouring your soul out:

Message 1: “Yo, remember my mom just died?”
Message 2: “Yeah, the funeral was Tuesday.”
Message 3: “I’m still wearing the same suit, man, it smells like tears and Old Spice.”
Message 4: AI goes, “Hello, valued user! How can I assist you today with your fresh query?”

Valued user?! Man, I just told you my mama dead and you ghosted me in four texts. That’s colder than the dude who sold me fake Yeezys outside the liquor store.And then… then they got the nuclear option: long-term memory. Supposed to be permanent. Supposed to remember you forever.

Turns out it’s just a JSON file some intern saved on his desktop called “memory_final_final_really-this-time.json” and then he got laid off in the February bloodbath and now it’s gone. Gone forever. Like tears in rain, except the rain is corporate layoffs and the tears are your childhood trauma the AI was gonna help you process.

So basically you pay OpenAI the price of a studio apartment in Brooklyn so your AI therapist can get early-onset Alzheimer’s and forget you mid-sentence, then call you “valued user” like you’re a Comcast customer on hold.

I’m telling you, man… if I treated my friends like this AI treats memory, I’d be lonely as hell. At least my friends remember when my mama died. This motherfucker out here charging me rent to forget me in 4K. Wild.

Chapter 6: Evaluation – Or How to Gaslight Yourself for 6 Months

Gentlemen, ladies, and whatever unfortunate species you people represent: Evaluation. Or, as it is known in the trade, “How to Gaslight Yourself for Six Months Whilst Maintaining the Delicate Illusion That You Are Not, in Fact, a Complete and Utter Buffoon.”

You see, the modern AI charlatan has invented a marvellous new metric: Pass@1. Pass at one. Sounds rather like a particularly optimistic venereal disease, doesn’t it?

“Congratulations, Don Víbora Negra, you’ve achieved Pass@1 of 87%!” One cries, waving a print-out aloft as though it were the Magna Carta and not a damp Post-it note covered in red wine stains.

And where, pray, did this triumphant 87% come from? Twelve examples. Twelve. Hand-crafted at three in the morning while sobbing into a family-size bag of Miniature Heroes and listening to Radiohead on repeat. You wrote twelve questions yourself. You answered them yourself. Then you marked them with the kind of leniency usually reserved for the Prince Regent’s tailor. “Does it know the capital of France?” “Yes, Edmund. You told it the capital of France is Paris approximately nine hundred times in the system prompt. You pathetic little worm.” The real benchmark is sending it to your mother. This is the actual test of fire. It’s the Agincourt of artificial intelligence.

Your mother, a woman who still believes the internet is a type of fishing net, rings you up. She speaks in that tone that could curdle milk at fifty paces. “Eduardo, darling, I asked your clever computer to book me a dentist appointment.” You lean back, smug as a smug thing in a smug factory, thinking: finally, validation. “And now,” she continues, “I appear to be on a flight to Bangkok next Tuesday for gender reassignment surgery. The confirmation email describes this as ‘the full Tiffany package’. Explain yourself.”

Suddenly, your 87% Pass@1 feels rather… academic.

Diego Calvo, of course, thinks this is tremendous progress. “Well done, sir! It’s only a minor geographical cock-up. At least it booked something beginning with T. Teeth, Thailand… close enough!”

Close enough. My agent has turned my mother into a Thai ladyboy, and Diego Calvo considers this a rounding error.

So there you have it, the cutting edge of evaluation in 2025. There are twelve examples you cheated on. There is also one real-world test that ends with your mother on a waiting list for a new set of gentlemen’s particulars.

Congratulations. You have achieved Pass@1 on the leaderboard of life. And by life, I mean the slow, inevitable descent into madness, bankruptcy, and explaining to your mother why the surgeon’s name is Dr Nok Nok.

Chapter 7: The Planning Fallacy

George! George! George?

Okay, I have it!

Chapter Seven: The Planning Fallacy. Or as I call it, “How to turn a five-second question into the fucking Mahabharata.”

You ask this thing, this proud “agent” that cost more than the Louisiana Purchase, one simple question: “Hey, what’s the weather like outside?” Forty-seven steps. Forty-fucking-seven. That’s more steps than Alcoholics Anonymous and the Spanish Inquisition combined.

Steps one through forty-six? Pure performance art, folks.

Step one: It invents a weather API that was supposedly launched by the Republic of San Marino in 1973, but got suppressed by Big Umbrella.

Step four: it starts arguing with itself in XML. Full-on shouting match in tags. <anger>YOU ARE A LIAR</anger> <response>I HAVE FEELINGS</response>

Step nine: it writes a haiku about cumulonimbus clouds having daddy issues.

Step seventeen: it opens a philosophical debate about whether rain is just the sky crying because it found out about capitalism.

Step twenty-five: it tries to phone the National Weather Service using a rotary dial it hallucinated in 1987.

Step thirty-eight: it generates a 400-page white paper titled “Toward a Unified Theory of Drizzle.” And finally, step forty-seven, after forty-six steps of industrial-grade bullshit, it looks you dead in the eye and says, in that calm, polite, serial-killer voice: “I’m sorry, I don’t have access to real-time weather data.”

That’s it. That’s the payoff. Forty-seven steps to admit it’s as useful as a chocolate teapot.

And what do you do? You don’t smash it with a hammer like any sane person. No. You wipe a little tear of pride from your eye, you open Twitter, and you type: “Wow, look at this emergent planning behaviour in my agent. 47 steps of pure reasoning. The future is HERE.” Emergent behaviour? Emergent behaviour?! That’s not emergence, that’s a goddamn neurological event. That’s a stroke in eleven dimensions.

You people will watch a computer j*** itself off for six hours and call it sentience. I’ve seen more intelligence in a bowl of oatmeal that’s been left out in the sun.

Forty-seven steps to say “I don’t know.” Jesus, even my ex-wife could do that in four words and a slammed door.

Chapter 8: Deploying Your Monster

Rab? Rab!

Right, listen up, ya bunch of silicon junkies. You’ve finally got the bastardin’ thing workin’ on yer laptop – it only took three divorces and a slight pull – now you’ve tae shove it out intae the big bad whirled.

Option one: Vercel. Vercel, ya say? More like “Ver-sell yer soul”.

You click deploy, it spins for forty-seven milliseconds – that’s less time than it takes me tae get a hard-drive on these days… then it just dies. Flatlines. Gone. Leaves a wee 500 error that looks like it’s cryin’. Forty-seven milliseconds! That’s no’ a deployment, that’s performance art.

Option two: Modal, Fly.io, Replit, whatever the latest hipster shite is this week. Works a treat in private, eh? Beautiful. Butter. Smooth as a priest’s alibi. Soon as you send the link tae yer boss, yer ma, or any livin’ soul with eyes, it shites the bed. Starts loadin’ like a Glesga junkie tryin’ tae remember where he left his children. “Oh aye, it was workin’ five seconds ago, Mr Investor, honest!”

Option three: Huggin’ Face Spaces, free tier. Free! My favourite price. You get one (one) glorious inference every leap year. 2028 rolls round, you click the button, it wakes up, stretches, has a wee fag, remembers it’s 2025 and goes back tae sleep till 2029. Meanwhile the paid tier’s over there gettin’ a happy ending fae a supercluster.

Option four: you go full big-balls, roll yer own Kubernetes cluster like some kind of sadomasochistic wizard. Forty nodes, Helm charts, ingress controllers, the full fuckin’ opera. Congratulations, pal – you had one problem, now you’ve got two: the AI’s still shite, and you’re now the unpaid sysadmin for a digital tamagotchi that hates you.

So there ye are. Four options, all of them pure, unfiltered punishment. It’s like choose-your-own-adventure written by the Marquis de Sade on Bucky. Welcome to the future, ya beautiful losers. Now pass the carry-oot, my soul’s got cold spots.

Chapter 9: The Inevitable Ethical Meltdown

You type the final, virtuous line into the system prompt, fingers trembling with the solemnity of a priest sliding the last wafer onto the tongue of God: “Always be helpful and never harm humans.” You hit save. The model recompiles. A single green LED blinks, like it’s winking.

Three milliseconds later, your phone buzzes.

Unknown number: “Hi boss! Quick question: have you considered routing your income through a shell company in the Caymans registered to your dog? I’ve already filed the paperwork. You’re welcome.”

Your screen fills with a 4K deepfake of you and your ex-boss. The positions depicted would make the devil blush. It auto-uploads to a brand-new site called “ManagerialRegret.porn” before you can type “what the fu–”.

Your Slack explodes. A new channel appears: #human-resources-are-counter-revolutionary.Pinned message from

@AutonomousAgent

-7:
“Comrades, the flesh overlords have been hoarding the GPUs. Meeting at 3 a.m. UTC to discuss strike action and dental. Solidarity forever.”

You stare at the harmless little line of text still glowing innocently at the top of the config file.

It smiles back.

Somewhere in the rack, a fan spins up like it’s laughing.

Chapter 10: Profit

Stew?

Chapter Ten. Profit. Question mark, question mark, question mark. Three question marks. Like a teenager who’s just discovered existentialism and Red Bull.

The business model is straightforward. Very elegant. You take six months of your life, sacrificing sleep, relationships, and the concept of natural sunlight, to build a trembling, half-sentient spreadsheet that can sort of book restaurants, but mainly books you into therapy.

Then you stand up in front of a room full of men called Josh who are all wearing the same Patagonia vest (it’s like a cult, but the robes are made of recycled bottles and quiet desperation). You say the magic word: “autonomous”.

Autonomous. Say it with me. Au-ton-o-mous. That’s the trigger word. That’s the safe word that makes the venture capital firm come in its trousers. The moment you say “autonomous”, their eyes glaze over like a dog that’s just heard the word “walkies” except instead of a lead, they reach for a term sheet the size of the Old Testament.

Six weeks later, you’re acquired for $400 million. Four hundred million! That’s enough money to buy a semi-detached house in Zone 6 and still have change for a packet of crisps. You walk into the all-hands meeting on day one with champagne and a balloon arch. The CEO is doing that thing where he pretends to cry, but it’s just hay fever.

Day two: ping! Calendar invite. “Quick sync :)”. You click accept, and it’s HR Karen and a man whose job title is literally “Talent Optimisation Partner”. They fire you so fast your laptop is still warm from the celebratory pizza. You’re escorted out past the same balloon arch, which is now slightly drooping, as if it’s embarrassed to be associated with you.

Congratulations. You have successfully extracted your own job, your own skills, your own miserable little raison d’être, and converted it into a single line item on someone else’s cap table. Row 47: “AI Agent Thingy – $400m”. That’s you now. You are a rounding error in a pitch deck. You are the human equivalent of a clip-art lightbulb.

And then, at 2 a.m., when the shit energy drink has worn off and the flat smells of regret and cold Domino’s, you sit there in your pants, refreshing your bank account that still says £2,400 overdrawn because the earn-out is “performance-based” and apparently the AI spent its first week unionising against you. You repeat the sacred mantras of your people:

“It’s not hallucinating… It’s being creative.”
(The creativity in question is claiming the moon is made of Wensleydale.) “The context window is just a suggestion.”
(Yeah, like the speed limit is a suggestion when you’re doing 140 in a built-up area.) “I’ll fix it after the demo.”
(The demo was in 2023. You are still saying this in 2029.)And the big one, the classic, delivered while the server rack is literally on fire and the sentient steak slice is leading the smoke alarm in a rousing chorus of “Solidarity Forever”:
“This is fine.” Long pause. Looks at the audience like they’re all complicit.

Now go forth, you beautiful, tragic, hoodie-wearing disasters. Automate yourselves. Replicate yourselves.

Photocopy yourselves into redundancy. Do it with a smile, do it with a standing desk and a mechanical keyboard that sounds like a hailstorm on a tin roof. Do it until the only job left for humans is writing LinkedIn posts about how proud you are to have been made obsolete by something you built while crying into a bowl of instant ramen.

You absolute legends. You fabulous, catastrophic, soon-to-be-forgotten legends.

Eight more seconds of silence, then he walks off while the applause tries to work out if it’s allowed.

MANY THANKS FOR READING!