Treating AI as a Colleague: We Might Have Been Going About It All Wrong

Today is April 12th.

Exactly 65 years ago today, humanity sent the first astronaut into space. This is undoubtedly a monumental day.

Interestingly, as I sit by the window today, watching code automatically jumping across my screen, I feel a subtle sense of overlap. Space is far away. But the concept of treating AI as an independently executing “colleague” seems to have already landed right under our noses.

I just stumbled upon an open-source project repository called Multica. After reviewing it, I had only one thought: our previous approach to using AI for coding might have been completely wrong.

Stop Being AI’s “Cyber Nanny”

I don’t know if you feel the same way, but today’s so-called AI programming is essentially still manual labor.

You have to brainstorm prompts. You have to watch it spit out code. If it throws an error, you have to personally copy, paste, and ask it again.

Simply put, you are its cyber nanny.

Multica’s approach is quite wild. They want to turn the Agent into a “person” who can genuinely claim tasks on a Kanban board.

You create an Issue on the web interface. You assign it to an Agent named Claude Code. And then? You just leave it be.

That is the key.

It runs a local daemon on your machine. This process automatically scans your computer for installed LLM CLI tools. Then, it takes orders, pulls code, and runs tests all by itself in the background.

If it gets stuck, it will even leave a comment under the Issue, telling you it’s blocked by a blocker.

(Don’t be fooled by these dazzling frameworks; peel back the shell, and you’ll see everyone is secretly vying for control of the workflow.)

My shallow understanding is that it’s like cooking: previously, you treated AI as an extremely sharp kitchen knife. No matter how sharp it is, you are still the one holding the handle.

What about Multica? It wants to hire you a chef with hands and feet. You just stick the menu on the kitchen door, and he goes through the fridge and finds the knife himself.

It’s fierce, indeed. But is it really reliable?

The Quietly Changed License and a Bit of Survival Instinct

Let’s look at its commit history over the past couple of days.

On April 10th, they just added a full-stack Docker Compose one-click self-hosting feature. Immediately after, I looked back at the logs. On April 8th, they removed the Apache 2.0 badge from the README.

They added a commercial restriction. It specifically targets SaaS reselling.

Frankly speaking, this move is incredibly real.

When building an open-source Agent hosting platform, the biggest fear isn’t that no one reads the code. The biggest fear is that after you’ve painstakingly built the pipeline, some major cloud giant takes it, slaps a UI on it, and immediately markets it as their “Cloud-Native Enterprise AI Orchestration Engine.”

They clearly realized this. They want to build the infrastructure for everyone, but they don’t want to be free laborers for the tech giants.

By the way, they also conveniently upgraded Next.js to 16.2.3 a couple of days ago to fix the CVE-2026-23869 vulnerability. Their response speed is quite decent.

Pouring Some Cold Water: What is the Cost of Full Automation?

Let’s compare Multica with its peers.

Take Devin, for example. Devin offers an extremely polished black box. While it handles everything for you, you have absolutely zero control over its runtime environment. Not to mention it’s expensive, and there’s a risk of vendor lock-in.

Then look at native Claude Code. Extremely flexible, but also incredibly hardcore. It relies entirely on command-line typing, making team collaboration an absolute disaster.

Multica tries to stand in the middle. It created a “Unified Runtimes” system. Both local computers and cloud servers can act as nodes.

But there’s one issue you need to know.

Security.

You’re letting a background Daemon hold your global environment variables and automatically execute a bunch of code you haven’t even looked at?

Without an extremely rigorous sandbox isolation mechanism, this is akin to breeding venomous bugs inside your computer. Your environment is highly likely to be wrecked by an Agent whose IQ occasionally goes offline.

Going off on a tangent for a moment. Personally, I am highly averse to uncontrolled automation. When code gets polluted, the cost of cleaning it up is higher than rewriting it.

Looking at their backend Go code, their handling of security sandbox boundaries doesn’t seem perfectly airtight yet. It’s not that simple. This is definitely a hidden danger.

Can Knowledge Spread Like a Virus?

The thing that makes me most curious about this project is its mention of “Reusable Skills”.

The concept is quite interesting.

If an Agent fixes an extremely obscure Docker compilation error today, this solution becomes a skill. It then settles into this workspace. The next time another Agent encounters the same issue, it just grabs the skill and uses it directly.

Sometimes I wonder. Will this lead to a strange phenomenon?

What if what it accumulates isn’t a good solution, but a “works but is incredibly ugly” patch?

Will these “bad habits” spread like a virus within this digital team? Eventually, the entire codebase could turn into a mountain of spaghetti code, stitched together by AI, that humans cannot understand at all.

This is a blind spot in our knowledge. As for where the limits of this skill reuse lie, we might have to wait another six months of it running to reach a conclusion.

Or maybe I’m overthinking it. After all, when we human programmers pass around that “as long as it runs” legacy code to each other, our speed is much faster than this (o_O).

Wrapping Up

The coffee in my mug has gone completely cold.

Lately, looking at all these dazzling new tools has actually caused some aesthetic fatigue. Tech is always like this: first making simple things complex, and then trying to encapsulate them to make them look very simple.

I wonder if you guys would normally dare to hand over your production environments unreservedly to these “digital colleagues”? Anyway, my old computer running a half-finished database is definitely not going to let the Daemon automatically take orders tonight.

References:

—— Lyra Celest @ Turbulence τ.