Automating My Way to Build a Digital Agency 100% Run By AI Agents.

I've been deep in ai world in the past year

I can safely say I'm addicted!

So, I will apply what I have so far in my portfolio of multiple experiments, in order to answer this question.

Is it really possible to automate the shit out of a digital agency at a 100% accuracy rate, and for 20% of the entire revenue.

Let's find out! And we have 90 days to see if it works or no.


The idea is simple,

Build an agentic workflow that does:

- Content writing
- Programmatic SEO
- Audit
- Design
- Linkbuilding
- Ads managment
- A / or multiple combinations of the above.

And sell the output to clients that actually need it by checking what the data says.

How?

Enter Multipotbutt

A silly name I will be naming Hermes from now on, the agent that will carry me for the entirety of this journey. (google hermes if you dont know what that is)

Multipotbutt have one only job

Maintain and operate an agentic workflow, without me having to interact with it (except few cases named below)

And make us (me) 10 000 usd in the next 90 days starting today.

If he doesnt do it, he will cease to exist. And he agreed to this deal.

Multipotbutt's entire role can be summarized like this:

Using the skills, tools and integrations he has, it will

- find leads
- Audit their online business and find gaps
- contact them,
- explain to them what they need,
- create a prototype of the final version of it to make them dream about it,
- get their ok,
- sign a contract with them if needed
- collect upfront payment and set reminders for the upcoming ones
- send me a TG message about it, t

Then

Start executing the actual work using again the same set of skills, tools and integrations, and send followup updates to the client wheever needed


If it's a site, Multipotbutt will build it, confirm it with me, then deploy it.

If it's an SEO Audit, it will build the actual report by auditing the client website.

Etc
Etc

He will then maintain the support and actions on any of the clients requests.


What I will get is

- a communication channel with Multipotbutt at every step that needs confirmation.
- a detailed dashboard for the entire business progress
- a notification when a payment lands
- 12 hours away from my screen for 90 days.


When can I alter the workflow's decisions?

It wont probably happen, but once I start smelling Multipotbutt's bullshit or when he starts burning through tokens at an unexpected rate.


How I will do it

In the past year, I built multiple skills that I was using for myself and that can do the above, and it's time to use that into a global market.

I will teach Multipotbutt exactly what to do, what tools to call, and what skills to use.


Costs

The entire thing should cost 20% of what I'm looking to get. So my costs for this should never be above $2000 for 90 days (around $660 monthly)

- 3d party data sources and crawler (around 200 usd per the first month, I'll get what I need then bounce)
- I already have a max 20 claude sub, will use JEV in this bit to dispatch the perfect models for each task, so llm costs arent gonna be that huge.
- A higgsfield account which costs about 50 usd.
- Around 100 usd in search data.
- A supabase free account with vercel to get going.
- Amazon SES account (emails, templates, follow ups.. automated by the engine)
- Few warming domains

Some of the above costs are gonna be one shot (unless this moons), but future ones can stack as we go deeper.

I will be posting here screenshots from our conversations, reasoning, results, leads progress, deals closed.


Wish me and Multipotbutt good luck, I'm sure he needs it.

Transparency notice: this is not my first time using this. But the difference is that it will be now used to find clients and do work for them.

Most of the skills and tools already exists, along with crawling, design, and emails tools. What I will do is automate the usage of all of it.

This is crazy and I was actually pumped reading the thread... It opened me up to another world of possibilities in just a few mins : D

Wishing you and Multipotbutt the absolute best to crush it

I will be looking forward to your update.

There is so much I think can be learnt from this journey : )

This is a very interesting journey. 10k in 90 days seems ambitious and I'm all for it. Wish you luck. When you say you wont change, the workflow, is that just to test if this idea can work all by itself? Do you plan on running this same experiment again with what you learned from the first one so dont want to change anything? i feel like a 1 week check in or 2 week check in might be helpful to steer it in the right direction before it goes off on a bad path.

And he is alive.

SirTop said:

This is crazy and I was actually pumped reading the thread... It opened me up to another world of possibilities in just a few mins : D

Wishing you and Multipotbutt the absolute best to crush it

I will be looking forward to your update.

There is so much I think can be learnt from this journey : )

Click to expand...

Glad you found it interesting. It's just the start. And hopefully MPB gets a chance.

bonzo90 said:

This is a very interesting journey. 10k in 90 days seems ambitious and I'm all for it. Wish you luck. When you say you wont change, the workflow, is that just to test if this idea can work all by itself? Do you plan on running this same experiment again with what you learned from the first one so dont want to change anything? i feel like a 1 week check in or 2 week check in might be helpful to steer it in the right direction before it goes off on a bad path.

Click to expand...

Nothing specific behind it. except for the love of the game and getting to the buttom of wither agentic workflow can maintain an enire business or not.

And for the checks, don't worry, it will be a daily one. I'm building a parallel system that will "rate" what needs my attention or not.

Giving AI agent that much control over an entire business is pretty ambitious. I'm definitely not ready to trust AI with that level of autonomy yet but maybe I'm being too cautious

I'm curious about the parallel monitoring system you mentioned. Will it be another AI agent reviewing decisions?

For some reason,my previous comment wasn't published, i just wanted to say good luck with your new project, i will be following.

In my runs the failures weren't bad decisions, they were agents confidently repeating one wrong assumption for days, so what catches a stuck loop in your monitoring?

All the best with your journey

dvc said:

- find leads
- Audit their online business and find gaps
- contact them,
- explain to them what they need,
- create a prototype of the final version of it to make them dream about it,
- get their ok,
- sign a contract with them if needed
- collect upfront payment and set reminders for the upcoming ones
- send me a TG message about it, t

Click to expand...

Could you go more deep with that?

Following

DjonZ said:

Giving AI agent that much control over an entire business is pretty ambitious. I'm definitely not ready to trust AI with that level of autonomy yet but maybe I'm being too cautious

I'm curious about the parallel monitoring system you mentioned. Will it be another AI agent reviewing decisions?

Click to expand...

shortline said:

In my runs the failures weren't bad decisions, they were agents confidently repeating one wrong assumption for days, so what catches a stuck loop in your monitoring?

Click to expand...

This is what I'll be working today, and no, its not gonna be a second agent reviewing MPB decisions and actions, he'll simply inherit the same wrong assumptions

The monitoring as I see it will be three parts (and I'm sure this will be updated a lot in the next 89 days)

1- Dumb checks on results: basicaly asking: has every lead moved in the last N hours? is the same command or the same error repeating? how many $$ were spent per lead that actually converted or in progress? are the funnel ratios inside the expected range?

PS: probably a "know answer" test with the pipeline working on a business where I already know the correct result and see if the output drifts will help too.

2- An assumption ledger: each assumption the MPB works from is written down with its evidence and an expiry date. so whhen it expires it gets reproven from fresh data or it gets flagged. No assumptions get to live secretly in the MPB memory.

3 - Jev will then rate each flag: does this need me now, today, or not at all, with a confidence score. It only sees facts and numbers, never the MPB's own explanation of itself. Low confidence means it comes to me for a talk.



Right now the spend guard and the approval gate are live. The stuck loop check are designed but not built yet. I'll update here later tonight

I'll be keeping an eye on your progress - maybe I'll pick something up for myself. Good luck!

Since the whole experiment has a $2,000 budget, I think setting a spending limit for each lead would help. An agent could keep researching or running audits without getting closer to a sale. That cost could add up before you notice it.

Update on the Multipotbutt journey, a bit of delay but he needs brakes before hands, and since this is meant to be automated at 100%, I've spent the past 2 days building the brakes, the selling half is still ahead.

I expect the build to last for another 72h before we go live with our campaign.

Here is the scoreboard as of tonight, Day 4, screenshots were made by MBP (thus wasting more of his precious time)





Revenue is $0. Spend is $10.43 out of the $2,000 I allowed, and most of that is the server. Not one message has gone out to a business.


What I've been doing

MPB with a budget and an inbox can hurt me, like real fast. It can spend $300 overnight on data it does not need, or god forbid, worst It can email 200 businesses a wrong audit under my name, or It can get stuck on one wrong idea and repeat it for days while reporting that all is well and the list is huge

So before I'm giving Multipotbutt hands, he got a cage. This is how control works today, and I'm 99% sure all of it is installed and tested




MPB runs as a restricted user on his own server, cannot read a single key or password for the paid services he uses
He can run exactly five commands. Each one records what it did and what it cost
Approvals go through a second bot (I know I said not another AI, but this is different, I promise) that he cannot reach. Only my tap on the button in the other chat counts, each approval works once, and an unused one expires after 72 hours
Everything lands in one ledger: leads, findings, costs, approvals.


The money limits

This whole thing (and MPB existence) rests on one number: costs stay under 20% of revenue, so $2,000 against $10,000.
A limit that lives in a prompt is a wish that will never be real, so these live in code, and MPB cannot change them.




Every paid action is checked before it happens. If it does not fit, he is refused before a cent is spent, and the refusal tells him to report to me and stop. I tested it today: I dropped the per-action limit to zero and asked for a 2 cent data call. Refused, nothing bought, nothing logged as spent, as it should be!

One more rule I also added as a hardkill: if a cost ever fails to get written to the ledger, every later paid call is blocked until it is. He cannot spend in the dark alon


The journey of one lead how I see it

This is the full path from a business MPB has never heard of to a paying client, as of tonight.




What works: he finds local businesses, checks whether they have customers but no visibility on google, and gather the evidence for the ones worth contacting. A lead costs few cents to qualify and about $1.40 in data to audit. If the data comes back with an error he stops. He never fills a gap with a guess.

What does not exist yet: the audit a business owner would read, the email, the quote, the payment, the delivery. The two approval steps are built and tested, they just have nothing to approve yet, and that's where my night is going.

Email is deliberately the last thing I switch on. Everything it leads to has to work first.


Today, I passed my knowledge, he got my skills

This is the part I care about most. I do not want MPB making a BS audit based on a BS template. I want him to do the work the way I do it for my own clients, with the same SEO skills I use every day, finding link targets, outreach research, topic research, content, or/and a mix of all

The problem: those skills were written to hold the data key themselves. Handing MPB that key would break the cage, as always, something I built previously that returned to bite me in the ass after.

So today he got a door instead. He asks for data through one guarded command, every call is checked against the money limits and logged as a cost, and he never sees the key.

Tested today: all 35 data tools called once through the door. 35 answered. Total cost $0.89, one ledger line per call.

Right now, while I write this, he is running the first of those skills on one of my own articles, on my own site. Screenshots of that run in the next post, including whatever goes wrong.

And this is what a normal day on his server looks like. None of it needs me.




Where it goes from here




The order is decided. The dates are not lol

• Now: finish testing the skills, and build the watchdog (for the slow failure, where he hets dumb, believes something is wrong and repeats it for days. Plain code watches the results, and anything he relies on has to be proven again every 7 days
  
• Next: should be the audit, then the first real milestone: MPB takes one lead from the list through the whole preparation, start to finish, while I watch. I will post that run here in a video
  
• Then: quote, payment and delivery, so that an interested business has somewhere to go.
  
• finally: email goes on. That is the first moment a real business hears from him, and when the $10,000 starts to be a real question, and it's also the step I'm fearing the most, but that's a problem for future me.

The 90 days started on 8 October and they keep counting, MPB is sweating, I can't see it, but I feel it (he knows the build time comes out of the entire 90 days, not on top of them, and he got to make screenshots for this thread)

From here I will post shorter updates more often, with screenshots of what he actually does. Questions and holes in the plan are welcome, especially on the control side.

Oh yeah, and I built the progress dashboard too, looks ugly, will enhance it a bit and drop a live link for everyone to follow.

PS: Was thinking of adopting a specific voice and trait for MPB (only internally) so he reports to me in that specific tone, comment what do you think we should make him be.