How Atlas decides which AI does the job.
When I say Atlas "decides" which AI to use, it can sound like magic. It isn't. It's a short list of rules, and I think you should be able to read them. I'm a student who built this on my own, so I couldn't afford rules I can't explain. Here's the list, and why each one is there.
Rule one: cost first
If a job can be done by something that doesn't cost me anything extra, it goes there. That means a local model running on my Mac, or a subscription I already pay for. The paid API is the last resort, because that's the one that charges per use.
I picked this rule first because I started out nervous about leaving an agent running on its own. The fear wasn't that it would be wrong. It was that it would be wrong expensively. So Atlas tracks what it spends and shows it, and there's a hard ceiling I set. When it hits the ceiling, it doesn't stop dead. It shifts toward the free options and keeps going.
Rule two: I can always override
If I pick a model myself, Atlas uses it. Always. The automatic choice is a default, not a decision I can't change. I added this after a few evenings of fighting a system that kept "helpfully" choosing something other than what I'd asked for. If I'm the one paying attention, I should win.
Rule three: don't send a tool job to something that can't use tools
Some models can't call tools. If a job needs to read a file or run something, Atlas won't send it to one of those. It's a small rule, but it fixed a whole category of confusing failures. Before it, a job would go to the cheapest model, which would politely describe how it would do the task instead of doing it. That looked like success in the log and wasn't.
Rule four: ask before anything sensitive
Each agent has its own level of trust, and some actions need approval no matter which agent asks. When you say "always allow," Atlas remembers only the narrow thing you approved, like one folder or one recipient. It doesn't quietly widen that into permission for everything.
This was the rule I argued with myself about most. Asking every time is annoying. But a tool that acts on my Mac should be a little annoying about the dangerous things. I'd rather click "allow" ten times than find out later what it did without asking.
Rule five: hand work down, bring a summary up
Work moves down through the agents, and each one returns a short digest, not its whole conversation. This is partly about cost, since passing full transcripts up a chain gets expensive fast. It's also about me. If I had to read everything every agent said, I wouldn't use the app.
How I check it works
I wrote more than 750 tests that run without internet and without spending anything, so I can change something and know quickly if I broke it. There's also a benchmark that checks the cheap default holds up against a stronger model on quality. If it doesn't hold up, the default is wrong, and I'd rather find that out myself.
What I'm still unsure about
Rule one is a bet. Cheap first is only a good idea if the cheap option is actually good enough, and that depends on the job. For some tasks it clearly isn't. I'd rather say so than pretend the default always wins. That's partly why it's an override and not a lock.
Overall
None of this is clever. It's a few honest rules, written down and tested. That's the point. I'd rather build something predictable that I can explain than something impressive that I can't. If one of these rules sounds wrong to you, tell me through the contact page. I read every message. You can join the early-access list at atlas-agents.ai.
Drafted and created with the help of AI.