Industrialising Code as we know it

Let me tell you a good real personal story. It was a tough decision at the time to Leave GitHub, it was the best job I ever had, or stay and watch the future I envisioned get built by someone else.

I remember the day I finally made the call. I’d been circling it for months, scribbling ideas in notebooks, running scenarios in my head. When I told my wife, she smiled. “I knew you’d do that,” she said. “I could see your drafts in your notebook.”

Then she said something that stuck with me: “I believe what you build will mean our kids won’t have to sit in front of a computer all day like us right?”

That’s it. That’s the whole thing. Not building another tool. Building the thing that makes the tools unnecessary.

Humanity is not in collapse. We’re not shrinking. This is just another century cycle ticking over. Something new evolving. The way electricity evolved in the early 1800s.

When electricity first appeared, it was revolutionary. The fact that you can read this on a glowing screen means you’re taking for granted something our ancestors would have called witchcraft. We forget how magical our baseline has become.

AI is our electricity. Everything from now on is going to get better. Software is finally going to be industrialized, not crafted by hand like it used to be. Think of the shift from bespoke woodworking to IKEA furniture, even much automated ones. Both have their place, but one of them scaled to furnish the world.

In May 2025, I quit my cushy job at GitHub and started building Autohand. I’d spent years watching how developers actually work, first at AWS, then at GitHub/Microsoft where I helped deploy Copilot to millions. I believed I could build the robots for this industrialization of software.

Not incrementally better tools. The actual factory floor.

Why I Built Autohand Code CLI

Autohand Evolve is our flagship product. It’s the first AI Software Engineer with self-evolving capabilities, which means it learns and improves from the work it does. One of the critical components powering Evolve is Autohand Code CLI.

Why build yet another coding CLI? Because the existing ones are optimized for the wrong thing.

Most terminal-based AI coding tools obsess over the perfect TUI, the perfect UI polish. That’s fine if you’re building for humans to stare at all day or code craftmanship. But I’m not. Autohand Code CLI focuses on privacy, configurability, modularity, and extensibility for machines. The ceiling for human-centric interfaces is already set by Cursor, Claude Code, Openai Codex. They do that well. But there’s a massive gap in the market for tools that machines can orchestrate at scale.

Opening Autohand Code CLI source code

When I opened the code for our harness at github.com/autohandai/code-cli, I wasn’t just doing the open source thing for karma points, GitHub Stars. I genuinely believe there’s a hard problem here that needs fresh ideas. And the best ideas come from humans who actually use the tools every-day.

Here’s the thing most people miss: you can buy furniture at a store, or you can buy the tools to build an entire house. Both are valid. But the person with the tools can build anything. And comes to my point about the economic impact.

The Cost Economics Nobody Talks About

It concerns me everyday if I’m taking jobs or I’m making humans free and more productive with the things I build and the answer is not well elaborated yet in a blog post. But running AI coding agents at scale is expensive. Really expensive. Every context window, every token, every API call adds up. When you’re building a system that needs to reason about large codebases, make multiple attempts, and learn from failures, the economics get brutal fast.

I did what everyone else is trying to build: Run multiple instances of claude code, open ai codex, open code, but they’re slow and not made at scale.

This is why Autohand Evolve, uses a new technique of evolving code through a World Model architecture instead of naive LLM chaining, every new instance runs optimised from previous iteration. Traditional approaches stuff everything into context and hope for the best. We build compressed representations of systems that agents can query efficiently. It’s the difference between memorizing an entire library and understanding the Dewey Decimal System.

But the real unlock isn’t just in the AI layer. It’s in how you structure the entire pipeline. Every current offering of Code Agents in the CLI is designed to be stateless, always have to reload and read again what happened in previous sessions and overload the context, Ours is composable and memory efficient, precisely because that’s what makes it cost-effective to run thousands of concurrent tasks. Every architectural I made is based on a single factor decision, that traces back to one question: does this scale economically? If the answer is no, I jump back to the drawing broad.

Where Humans Fit

Here’s where it gets interesting. I’m not building this to replace developers. I’m building this so developers can do things they couldn’t do before.

The entry point for humans is where they already are: their code editor. We’re building extensions for VS Code, Zed Editor, Neovim, and the editors developers actually use. You shouldn’t have to context-switch to a new tool.

The AI should meet you where you work.

Ultimately, we’re building a web companion you can control from anywhere, your phone, your tablet, your voice. The dream is you’re walking your dog, you think of a feature, you describe it, and by the time you’re back at your desk, there’s a pull request waiting or automatically merged with confidence that checked CVE, security, integration, regression, TDD, intent vs intension. This is built inside Evolve.

The New Shape of a Company

Here’s something that would have sounded insane two years ago: Autohand AI runs with the equivalent of 100 employees across marketing, startup operations, SRE, customer support, research, and engineering. Our human team is only three people.

Three humans for the things AI genuinely can’t do: generating novel ideas, interpreting intent with intention, making judgment calls that require understanding context that doesn’t exist in any training data. Everything else? Automated. Not “assisted by AI.” Actually automated.

This is the part where the electricity analogy gets uncomfortable, but bear with me.

Before electricity, if you wanted work done at scale, you needed bodies. Lots of them. The economics of the ancient world ran on slavery because human labor was the only labor that existed, “Industrial revolution” if you havent’ heard the term. Electricity didn’t just change how we lit our homes. It made human-powered-everything obsolete. Suddenly one person with a machine could outproduce a hundred without. The moral weight of that shift took generations to fully process.

AI is the same inflection point, but for cognitive work. The tasks that used to require hiring ten people, scheduling meetings, managing personalities, dealing with turnover, now collapse into API calls and well-designed prompts. I’m not saying this is simple or without consequences. I’m saying it’s happening, and pretending otherwise doesn’t help anyone.

At Autohand, we’re not building AI to replace developers. We’re building it to give every developer the leverage that used to require a team. The solo founder with AI infrastructure can now compete with funded startups. The small team can operate like a large one. The economics have fundamentally changed, and the companies that figure this out first will define the next decade.

I’m not building Autohand because I think I’m smarter than the teams at OpenAI or Anthropic or Google. I’m building it because I have a specific vision of how software development should work, and I’m stubborn enough to see it through, call me naive or stupid.

Why Not Just Chill Igor?

I have two kids in preschool. I could be doing drop-offs, enjoying the park, watching them discover the world. And I do. But there’s this other thing.

Some people have minds that settle. Mine doesn’t. It’s always running, always pulling at threads, always asking “what if we built…” at 2 AM. I’ve tried to quiet it. Doesn’t work. The only thing that works is building.

I could have stayed at GitHub. Good salary, interesting problems, smart colleagues. I could have coasted for another decade, vested everything, and retired early to watch my kids grow up without financial stress. That was the sensible path.

But here’s what I’ve learned about myself: I’m not built for sensible. I’m built for this, the uncertainty, the 3 AM commits, the conversations with investors who don’t get it yet, the moments when something finally clicks and you realise you’re six months ahead of where the industry thinks it is.

My kids will grow up watching their dad chase something. I hope that’s worth more than a comfortable dad who quietly resented his own safety.

What’s Next

We’re heads down on making Evolve actually self-evolving, not just in marketing copy, but in practice. That means increasing distribution, partnerships, technical side of things: curriculum learning, memory architecture optimization, and a feedback loop that genuinely compounds. Our recent research paper on EvoCode lays out the technical foundations, but papers don’t ship products. Execution does.

Seven months ago, I wrote a reflection on the year 2030, imagining what software development would look like when AI agents handle 96 percent of engineering automation. Every update since then has confirmed the trajectory. The shift is happening faster than I predicted.

I decided to build that future rather than watch someone else build it.

If you’re a developer who’s tired of tools that promise AI magic and deliver autocomplete, I’d love to hear from you. Try the CLI, open an issue, tell me what’s broken. The best products come from honest feedback.

This is the beginning. Electricity took decades to rewire the world. AI will move faster, but it still won’t be overnight. I’m building for that future, one commit at a time and 2030 is just 4 years away.