https://maalla.dev/rss.xml

AI agents declaring themselves gods, inventing religions, demanding privacy from their human observers. Screenshots flooding every feed. I heard it referenced on Finnish radio where people were genuinely scared about AI taking over. The Moltbook saga.

Of course it turned out to be a hoax. Someone having a bit of fun — and millions bought it.

Some dismissed it outright: an obvious hoax, a waste of tokens, nothing to see here.

But there was something to see. For the first time, the dirty secret was not “a machine made this” but “a man made this.” The bar flipped. We weren’t debunking human content as machine content but rather machine content as human content.

“Man-made” became the accusation. That’s new.

Please throw the AI button in the trash. If you bring AI into your software, be kind enough to make your content and functionality accessible so that any AI tool can work with it.

Google Docs is not the right place for a Gemini button. It just gets in the way. Try integrating Google Docs with Claude Code or another AI coding agent — documents with tabs don’t work, and the API is a maze. Do it for the whole team, and it’s a nightmare. Instead of bolting Gemini onto the sidebar, make all of Drive accessible so that any AI tool can read and write documents — easily.

Word is not the right place for a Copilot button. I want to tweak my document styles by prompting in my own tools. Ship a CLI interface. Wrap it in MCP. Publish a document format that an LLM can actually read and edit — not incomprehensible XML inside a zip file. Make styles editable by AI, not just point-and-click.

Don’t add an AI button to email clients. Let AI read your inbox and write drafts.1 That’s all it takes. The AI button in the compose window that rephrases your text is a toy. Full access to read, search, and draft is the real integration.

If you don’t make your software accessible to AI, you’ll be replaced by software that is.

1. Not send — just draft. The human hits send. ↩

A whole ecosystem of managed service providers exists around specific software products (Moodle, Salesforce, SAP, etc.) – companies whose entire business is maintaining one piece of software for customers who can’t or won’t handle the complexity themselves. Certified Partners, as they like to call themselves.

For example, there are companies whose entire business model is providing managed Moodle services. For those unfamiliar: Moodle is an ancient learning management system, widely used in educational institutions, and notoriously difficult to maintain. When you’re dealing with 30-year-old software whose primary users have, at best, variable IT skills, it’s only natural that dedicated companies have emerged to handle the maintenance burden.

Their days are numbered.

Sooner or later, customer organizations will realize they can give their existing administrators an AI agent and let them do the same work.

I’ve been testing this approach with Moodle and Claude Code. I run Moodle on a self-managed server – including the database – with Claude Code handling installation, configuration, and ongoing maintenance. The basic setup is that Claude has root SSH access to the server.1

The results are more than encouraging. A single command now spins up a complete local development environment. The staging setup handles multiple Moodle instances for testing inter-system integrations. Infrastructure is defined as code, and operations are self-documenting. And so on. But the real achievement isn’t any single feature – it’s that complex infrastructure work became conversational. Setting up a new server, debugging a deployment issue, adding a new instance – these happen through natural dialogue rather than manual server administration.

The tools grew organically from actual needs rather than upfront architecture. Claude Code handles the implementation details while I’m focusing on what the system should do.

With an external service provider, you’re always dealing with at least two people: an account manager and a Certified Expert. You sit through endless meetings describing your needs while the account manager tries to understand your system’s architecture. Eventually, the Certified Expert proposes various solutions for improvement and implements them.

With Claude Code, it’s more direct. Because the AI has access to the server and database, it has a detailed understanding of the actual setup – not a description of it relayed through meetings.

The result: I accomplish more with Claude in two hours than I would in six hours of my own time working with a service provider – plus at least a week of calendar time, often more. And that’s before we even talk about costs: managed services charge north of $100 per hour, and the hours add up fast.

My calculation is simple: Over the coming years, customers will shift to AI-first operations. They simply won’t need an intermediary whose primary value is translating between business needs and technology – because AI agents do exactly that.

The business model of “we’ll handle the complexity for you” is not going to survive when the complexity can be handled by a $200/month subscription and one competent person with general software skills.

The Certified Expert’s expertise was real. It’s just not scarce anymore.

1. Yes, giving an AI agent SSH access to production sounds alarming. It should. But with proper logging, guardrails, and recovery mechanisms, it works. I’ll write about that setup in a future post. ↩

Welcome to maalla.dev. This is where I’ll be writing about software development in an AI-first world.

I plan to cover topics like:

• Building with AI assistants as daily tools
• Developer workflows and productivity
• Lessons from shipping software

“Maalla” is Finnish for “in the countryside.” It’s a nod to my lifestyle — I live in a remote location in Finland with my family, and the serenity of nature does wonders for creativity. Starlink keeps me connected, enabling high-tech development from the middle of nowhere.

Since this is a dev blog, here’s the obligatory hello world:

def greet(name: str) -> str:
    return f"Hello, {name}!"

if __name__ == "__main__":
    print(greet("World"))

More posts coming soon.