Coding Agents are the Factorio for the Real World

The whole is greater than the sum of its parts.

That’s been true for teams. For companies. For organisms. Software multiplied leverage, but it didn’t multiply judgment. You had libraries. Frameworks. APIs. They helped. But they didn’t decide.

Now they do.

An LLM on its own is impressive. An agent on its own is useful. A loop: observe, call something, parse, act, repeat. Agents have existed for decades; they used to call heuristics, rule engines, policy networks.

Now the loop calls an LLM. And it’s wired to sub-agents that write code, search docs, run tests, deploy.

The breakthrough isn’t any single part. It’s the orchestration: routing, feedback, verification, and boundaries.

The wiring is the product.

The agent is the program. The LLM is the engine. Together, wired to sub-agents, they become something else entirely.

Before ChatGPT, there was CLIP. Give it an image and a set of text descriptions. It picks the best match. Zero-shot, no task-specific fine-tuning. That was the first wow.

The ChatGPT moment was different for everyone. For some it was writing an essay. For others it was debugging code. For me, it was search and deep search. A machine that could search the web, dig deep across sources, reason about what it found, and talk back. Not four separate tools. One conversation.

One window, all the answers. The juggling stops.

Coding agents are on the path to becoming the Factorio for the real world.

If you’ve played Factorio, you know the feeling: you start by hand-mining ore, smelt it in a furnace, lay your first belts. Then you automate mining and smelting, then the parts that build automation. Before long, your factory is launching rockets while you sip coffee.[2]

That’s what’s happening with coding agents. Hold that thought. We’ll circle back.

Same person. Different era.

Anyone with access to one will have a shot at building and providing value to the world. Online for now. Limited to software but people are already experimenting with adding lab in the loop.

But that itself is a huge opportunity.

A single-person unicorn startup.[1]

Let that sink in.

Let that unicorn in.

Up until now, this was incomprehensible to me. An app? An image editing app? A networking app? Maybe. But as you scale you end up needing more people.

But now, one can dream. An army of sub-agents working cohesively to achieve the goal. People have already prototyped this:

Devin (autonomous software engineer)
OpenClaw (parallel sub-agents, isolated contexts)
Claude Code with sub-agents

The agent factory. One person at the center orchestrating it all.

I’ve seen the before and after of the internet. The before and after of the iPhone. The before and after of ChatGPT (yes I am ~~old~~ oldish 😂).

Each one rewired everything. This feels like the next reset.

But let’s not kid ourselves.

The hype is real. Not everything will work. Most agent startups will fail the way most startups fail. Half of them are the torch and fart apps of the agent era. The tooling is fragile. The costs are high. The hallucinations are still there. And an agent with your credentials is a security surface nobody’s fully figured out yet.

Besides, more often you’ll use the extra leverage to do more, and end up just as busy.

Same person. Different era. More automation, more work.

And yet.

The ceiling has moved. What one person can build today was a team’s job last year.

Probably the greatest time of our lives.

So are you ready, player one? And yet.

Addendum: Agents before LLMs

The word “agent” isn’t new in AI. Long before LLMs, reinforcement learning agents were learning to play games, balance poles, and land a lunar lander in simulation. Sutton and Barto wrote the book on it 😉

Some explorations from 2018:

Solving CartPole with Hill Climbing - simple policy-based methods
Deep Q Network for Lunar Lander - based on the DQN paper that kicked off the deep-RL wave
DDPG for Continuous Control - when actions aren’t just left/right
Multi-Agent DDPG - two agents learning to cooperate

In 2023, Stanford’s Generative Agents paper showed 25 LLM-powered agents living in a virtual town, autonomously organising a Valentine’s Day party. Observe. Plan. Reflect. Act. The loop is the same whether you’re running a DQN or orchestrating a coding agent.

The RL agent ran a learned policy. Reward signals. Millions of episodes.

The coding agent calls an LLM. Context. One conversation.

Different engines. Same loop: observe, decide, act, repeat.

Notes

[1] Not horses selling courses, crypto, and YouTubers pointing at nothing with their mouth open. Think microfinance, child education, mental health, clean water, open science. If one person can now build what teams couldn’t, point it at something that matters.

[2] Massively simplified. Real Factorio players, please don’t @ me.

#CodingAgents #Factorio #SinglePersonUnicorn #ReadyPlayerOne #AgenticAI