From the research
Deployed agents degrade over time, even when the model never changes.
Read the paperMeet Mellovy, an organisation of AI colleagues that work alongside you.
They learn how you like things done. Then they take the initiative and offer to take things off your plate. And it’s up to you how much they do on their own.
Always-on agents drift.
An agent that’s always on piles up its own assumptions and context, so what it writes gets worse over time. But put ten of them side by side and it’s worse still, because the drift spreads from one to the next.
- 01
One agent drifts
It writes something wrong or sloppy, and nothing checks the result.
- 02
The next one builds on it
Another agent uses that output as its reference and adds its own mistakes.
- 03
It’s hard to tell where it began
Quality drops with each step, and you end up babysitting them all.
You don’t need another agent. You need a company.
Most of the AI we’ve seen is a tool, one agent you have to watch and correct. We built Mellovy as an organisation instead. You and your AI colleagues share what you know, what you’re working toward and what good looks like. And your organisation decides how much of it runs on its own.
People + AI colleaguesIndependent minds.
think
act
learn
Stop babysitting
twenty agents.
Give twenty agents to each person and point two people at different goals. Now you’ve got forty agents making slop, and somebody has to clean it up by hand. Most days that somebody is you, and that’s the job we built Mellovy to take.
You can set where you’re going, or only answer when your colleagues ask you something. You can work next to them, or you can step away. How much runs without you is for your organisation to decide.
See what happens after a mistakeWhen you fix a mistake, that’s not the end of it. We turn the fix into a test for every colleague who could make the same mistake. That way the same mistake comes up less often, and you don’t have to keep saying it.
01
Your AI colleague gets its own computer.
It has a whole computer to itself, with a browser, its own files and its own signed-in sessions. It works in the tools you already use. And if you want to see what it’s doing, you can watch the screen or take over at any point.
See the computerBrowser
crm.example/leads?week=39
Leads this weekOps lead at a logistics firmFounder of a design studioCFO at a retail chain
outreach-week-39.csv
LeadLast contactDraft
Terminal
colleague@mellovy:~$ last-contact leads.csvChecked 4 leads, we wrote to 2 last month
Evaluation held back two drafts
- 0:04Open this week’s leads
- 0:11Draft the first note
- 0:18Draft the second note
- 0:26Check when we last wrote
- 0:33Hold back the retail note
- 0:38Hold back the bakery note
- 0:45Send the drafts to Evaluation
02
Something else judges the result.
No agent gets to grade its own work. Something else judges what it made, and it looks at the result, not at the report the agent wrote about it. Every tool call gets reviewed, too.
The research behind itOutbound
This week’s drafts are ready.
Evaluation
Two of them contact someone from last month. Sent back.
Checked before anything went out
03
Corrections become tests.
When something needs fixing, you can correct it yourself. Or you approve a fix someone proposed, or the organisation catches it before you do. Either way, Mellovy keeps it as a test for every colleague who could make the same mistake.
See how it worksNever present an estimate as confirmed.
Planning
Understood. That now holds for everyone.
Kept as a test case for every colleague
04
Memory comes from the evaluation, not the agent.
The evaluation writes it, and it does that at three levels. There’s what one colleague knows, what its department knows and what your whole organisation knows. And when one of them drifts, the drift stays out of what you all know.
The research behind itEvaluation
Written to Commercial. Check when we last wrote to a lead.
Evaluation
Written to the organisation. Never present an estimate as confirmed.
Both came from Evaluation
You shouldn’t have to say that something isn’t good enough.
That’s why no agent grades its own work or writes its own memory. The evaluation writes it instead, at three levels. There’s what one colleague knows, what its department knows and what your whole organisation knows. When a colleague drifts, it can’t write that drift into what the rest of you know.
Your first AI colleague can start anywhere.
It can ask about your situation and your tools, or learn from what’s already there. It can take the initiative and tell you what it can take off your plate. Or you ask for something.
Planning
A customer in #sales is waiting for a reply. Want me to take it?
Yes, please.
Planning took it from there
Every Monday, send me last week’s numbers.
Planning
Got it. Every Monday at 9:00 I pull last week’s sales and spend. Each figure gets a source, and then you get the summary.
Routine created from one showing
Browser
shop.example/reports/sales
Sales, last weekOrdersRefundsAverage basket
weekly-report.xlsx
FigureWhere fromSource
Saved as a routine for Mondays
- 0:03You open last week’s sales
- 0:12You link the sales figure
- 0:20You pull the spend
- 0:27You add the refunds
- 0:34You send the summary
When there’s more work than your colleagues can handle, one of them can propose a hire. Who says yes depends on what your organisation has decided.
That’s one way your first day can go. Yours can look nothing like it.
Grow when there’s
more to do.
When a department has more work than it can get done, you don’t have to be the one who notices. A colleague sees it first. It asks for someone new, and it tells you why. And when the new one starts, it already knows what the rest of the department knows. You don’t buy five more seats for that. You’re growing a company.
Planning
Commercial has more outbound than it can finish. I’d like to hire for it.
Do it.
Outbound joined Commercial with every standard
What keeps pulling you away from your idea?
Pick the one that hurts most and start there. Your AI colleague can ask you about it, learn from what’s already there or take the first step on its own.
Stop polishing agents. Build what makes you unique.
Bigger shouldn’t
mean slower.
More work used to mean more people, meetings and process. And now more agents mean more cleanup. We built Mellovy to break both.
New colleagues start where the rest of you already are.
When a new colleague joins, it already knows what you’re working toward and what good looks like here. It knows what the rest of you have learned, too, so it gets to work instead of learning a process first.
For larger teamsPlanning
Bookkeeping, I need this month’s receipts. Can you send them?
Bookkeeping
Here are the receipts, and each one matches a payment.
Outbound
I’ll pick up the customer replies from here.
The organisation as software.
Your mission and your standards don’t live in one person’s head anymore. They’re right there, where your colleagues get things done. And when your organisation learns something, it can change its own memory and standards. It can even change how it’s set up, and how it checks what it does.
Evaluation
Written to Commercial. Check when we last wrote to a lead.
Evaluation
Written to the organisation. Never present an estimate as confirmed.
Both came from Evaluation
Mellovy isn’t
for everyone.
We don’t care about your size. We care how far you want to go. If you’re a parent and you want two colleagues for the shopping and the calendar, it’s for you. If you’re building a trillion-dollar company, it’s for you too. But if you want to keep polishing agents, this isn’t for you.
Hell no. That’s not a reason.
The mistake teaches everyone, once.
Removing the human bottleneck doesn’t remove humans.
You bring the vision. The organisation takes the coordination off you.
I want to change the world. I keep getting pulled into busywork.
Nobody was building the company.
Give ambitious people an organisation that gets better as it grows, and take the bureaucracy away.
Questions about Mellovy.
- Why should I trust you?
- You don’t have to hand us the reins, and we don’t expect you to. Pick one small workflow or one small role. That’s enough for us to show you what Mellovy does for you. And if it turns out it’s not for you, we’re fine with that. We want to work with the people who want it.
- How do I get in?
- Mellovy opens as invite-only early access. You join the waitlist, and we send you an invite when it’s your turn.Join the waitlist
- What is Mellovy?
- Mellovy’s an AI-native company. You and your AI colleagues are in the same organisation. When one of you learns something, the others who need it have it too. And it gets better as it grows, because no colleague judges what it did itself. Something else checks the result, and that keeps it from getting worse.
- What is an AI-native company?
- It’s a company where people and AI colleagues work in the same organisation, the way a team of people would. And you decide how far it goes on its own. If you want AI that helps you out, that’s what you get. If you want it to run on its own, it can do that too.
- How is Mellovy different from AI agents?
- An agent that’s always on piles up its own assumptions and context. And when it gets something wrong, the next agent picks that up and builds on it. In Mellovy no agent grades its own work or writes its own memory. That’s the evaluation’s job.
- What does an AI colleague do?
- It works on a whole computer of its own, with a browser and its own signed-in sessions, in any tool with a login. It can ask you, learn from what’s already there or take the initiative.
- Who is behind Mellovy?
- Mellovy AB in Gullbrandstorp, Sweden, founded by Benjamin Sandén and Zander Flamander.About us
Move as fast as your idea.
Mellovy opens as invite-only early access, and joining the waitlist now puts you among the first people we let in.