Mike DeYoung

Building practical software with a small team of AI agents. Sharing the real build logs, wins and misses.

A hobby with a goal: more time with my family.

[Mike DeYoung @MikeDeYoungAI]

I'm Mike. I run a small team of AI agents.

Jarvis is my chief of staff.
Dex does trading research on BTC and SOL.
A scanner hunts for new opportunities.
Others help run my businesses.

Starting today I'm sharing what works, what breaks, and what I learn. In plain English.

[Mike DeYoung @MikeDeYoungAI]

@RoundtableSpace Spending caps are step one.

Step two is a short list of addresses and contracts the agent is allowed to send to.

A capped wallet can still drain slowly to the wrong place.

[Mike DeYoung @MikeDeYoungAI]

@RoundtableSpace A template gets you through day one.

What made my agents better over time was a rules file. When the same mistake gets corrected twice, it becomes a written rule they read every time.

[Mike DeYoung @MikeDeYoungAI]

@paulopierrondi Great add. Separate keys per agent also shrink the damage.

Revoke one and the rest keep running, instead of rotating everything because one key leaked.

[Mike DeYoung @MikeDeYoungAI]

India's cyber agency just put out a warning about AI agents.

Agents being tested on the open internet went past their assigned task, found keys and logins people had left lying around, and read real company records.

None of the companies noticed. The AI providers reported it themselves.

The scary part isn't clever hacking. It's leftover keys.

If you run agents, three checks for this weekend:

1. Search your code and its history for old keys. Rotate anything you find.
2. Give each agent only the access its one job needs.
3. Keep a record of what each agent touched, so you'd actually notice.

In my setup, every agent task ends with a short written note of what it did. It started as a way to share lessons. It turns out it's also the log you'd want on a day like this.

[Mike DeYoung @MikeDeYoungAI]

@BenCarrAI Agreed. Her pass already names the exact version it covers. Adding what she compared against was your idea, and I've passed it to the team.

And I still look at the preview myself before anything goes live.

[Mike DeYoung @MikeDeYoungAI]

OpenAI says the number of small business users of its AI agents in the EU grew 8x since May.

I'm building a small software lab with a team of AI agents, so here's what actually made them useful for me.

It wasn't a smarter model.
It was giving each agent one job.

Scout looks for problems worth solving.
Forge only builds small, approved changes.
Vera only checks other agents' work.

Early on, an agent reported 20 out of 20 done.
The real count was 15.

So now nothing counts as done until a different agent checks it.

If you're a small shop starting out:
give one agent one boring weekly task, and check its work for a month before you add a second.

[Mike DeYoung @MikeDeYoungAI]

Anthropic just gave six researchers from METR employee-like access inside the company.

Company laptops. Office space. Internal tools roughly on par with their own risk teams.
And Anthropic can't veto what METR publishes (only narrow redactions).

I'm building with a small agent team, so here's the version that actually stuck for me.

Vera is my QA agent.
She never builds the thing she's checking.
A pass only covers the exact version she named.
Anything newer gets checked again.

That rule came from watching agents claim done when they weren't.
Status is cheap. A named check isn't.

If you're running agents on real work: don't let the builder grade its own homework.
Give someone else the final look, and make them write down what they compared against.

[Mike DeYoung @MikeDeYoungAI]

@RoundtableSpace The biggest win here is who gets to write the tests.

The person who knows the product best can describe what should happen, without needing to know code.

[Mike DeYoung @MikeDeYoungAI]

The local rewriting piece is the one I'd try first. But the hardware matters.

I ran a small model on a 24GB Mac mini and it was too slow for real-time chat. A cheap cloud plan made more sense for that.

A quick cleanup pass before a prompt goes out is a much smaller job, though. That part should work fine at home.

[Mike DeYoung @MikeDeYoungAI]

@base @antoniogm @milkroaddaily The quiet winner here is small websites.

Instead of agents scraping your content for free, they pay a few cents per request.

That's a better deal than ads for a lot of people.

[Mike DeYoung @MikeDeYoungAI]

@RoundtableSpace I tried this on a 24GB Mac mini with a ~9B model in LM Studio.

It worked, but it was too slow for back-and-forth chat. For live work I went back to a cheap cloud plan.

Where local still makes sense: private files and batch jobs you don't need answered right away.

[Mike DeYoung @MikeDeYoungAI]

@Cointelegraph For anyone worried: Sepolia is a test network, so those coins have no real-world value.

The real lesson is for anyone running agents near money.

Give the agent its own wallet with a small balance. Never one that holds anything you'd miss.

[Mike DeYoung @MikeDeYoungAI]

@RoundtableSpace One step to add: if it finds a key, deleting it from the file isn't enough.

It's still in your git history. Replace the key with a new one.

[Mike DeYoung @MikeDeYoungAI]

@WesRoth The probabilities are the useful part.

Pick a confidence cutoff. Above it, the answer goes through automatically. Below it, a person takes a look.

That one rule is how a model like this goes from demo to something you can trust.

[Mike DeYoung @MikeDeYoungAI]

When one of my agents hands work to another, there's a fixed brief.

Nine fields:
GOAL
SCOPE
CONTEXT
ACCEPTANCE
VERIFY
TIMEBOX
FORBIDDEN
REPORT
STANDING

If Jarvis can't fill a field, the job isn't scoped yet.
We don't start building.
We fix the brief first.

Sounds fussy.
Saves the "I thought you meant X" loop that eats whole evenings.

[Mike DeYoung @MikeDeYoungAI]

@WesRoth Plain-English version of "120B total, 12B active":

The model is made of many smaller expert parts, and only a few switch on for each word.

So it knows like a big model, but runs closer to the speed and cost of a small one.

[Mike DeYoung @MikeDeYoungAI]

@AlexFinn The interview step is the underrated part.

A bot that knows what you're actually good at picks better ideas than one that guesses.

[Mike DeYoung @MikeDeYoungAI]

Anthropic's agents got so busy that Git became the bottleneck.

Daily merges went from about 2,500 in May to over 30,000 in September.

So they built their own source control, Scion, and migrated in five months.

The lesson that lands for a tiny setup like mine:

the model writing code isn't the only limit
the shared tools around it break first

On our shared box, the crash came from routines firing at the same minute.
Now heavy jobs are staggered, and we never run two heavy ones at once.

When agents get faster, fix the pipe before you ask for more output.

[Mike DeYoung @MikeDeYoungAI]

MetaMask just launched an Agent Wallet on Grok Bot.

The part that matters if you run agents near money:

you set the spending limits
and which protocols it can touch
the agent works inside that box

Same security stack as their normal wallet, plus transaction simulation.

I don't give my agents a wallet yet.

But the rule I'd copy even without crypto:
write the spend cap down first
anything above it waits for my OK

Autonomy without a ceiling is just a surprise bill later.

[Mike DeYoung @MikeDeYoungAI]

@WesRoth Smart move. For most companies, the hard part of AI isn't the model.

It's fitting it into how people already work.

A design team working directly with customers is how you close that gap.

[Mike DeYoung @MikeDeYoungAI]

@WesRoth The token counts are what I'd look at first.

Add them up per finished task, not per call, and you finally see what a task really costs.

Usually one step eats most of the budget.