I Gave My Agents My Inbox, and They Made Me $220,000 [Part 1 of 3]

I had been running OpenClaw for a few weeks, doing things on Slack, starting to work on my email. It was genuinely fun. You throw a task at one of these agentic frameworks and watch it figure the thing out. But if I had told it something in the last couple of days it remembered, and as soon as a week passed I was teaching it the same random facts again. That is a frustrating way to work with something you are supposed to be handing work to. It is also, I think, most people's real complaint about ChatGPT and Claude. It just doesn't stick.

This is the first of three posts about what I did about that.

The thing that kept breaking

The harness was okay. The memory was not, and I had been chewing on why since Opus 4.5 landed and I first went down this road.

Around the same time I tried Paperclip, which takes a different shape. You start a call, and instead of one agent you get several. It generates a marketer, a growth person, and so on. The skills and the knowledge get assigned to different parties, so there is orchestration involved and no single agent has to know everything.

That was the piece that stuck with me. What I wanted was not a better agent, it was a structure.

Cascade

So I started building my own, and I called it Tartare. Yes, after a cut of meat again. I've explained why I do that before, and I'm told I am going to run out of cuts eventually. I already have.

The idea I kept coming back to is cascade. The head agent defines the tone and the language generally. Underneath it you have specialists that inherit some of that knowledge, some of the configuration, some of the skills, and then bring their own on top. It is an org chart, not a chatbot.

At the top of mine is me, the human. Then there are direct reports. Then there are categories: personal, family, each of the companies. Each agent has its own skills and knowledge files depending on what it actually needs to do.

Two of them I talk to directly. Thea is my counterparty for almost everything. I speak to her and she finds the right agent for the job. Tommy handles my Tomahawk fund, so every new investment opportunity goes to him, and he discusses it with his team and comes back with an answer.

They are not just processes. They have email accounts, Drive accounts, calendars. They are getting phone numbers now, so they can call people and confirm restaurant reservations.

The clearest example of the shape is my inbox. I use Missive, and I have for years. Originally I used it with Bianca, who I work with. An email would come in and I could say, Bianca, can you have a look, can you draft a reply. She would tell me what was urgent, what could wait a week, and quietly take a lot off my plate.

Thea has that role now. We did a lot of backtesting to get it right, and she genuinely understands which emails need to reach me today and which do not. She also extracts the attachments, names them according to our filing convention, puts them on Drive, and if it's an invoice she pulls out the amount and puts it in a spreadsheet so Bianca can double-check it.

Three moments where it stopped being a toy

The first was an investment pitch.

Tommy and Thea now know who I actually have a relationship with, based on past emails, WhatsApps, call recordings and my calendar. They have a good sense of who really knows me and who just found my email somewhere.

So an opportunity comes in. Thea decides it belongs with Tommy because it's an investment pitch and doesn't need a reply from me personally. Tommy picks it up, researches it, works out why it's a pass, creates a CRM card in Notion for each company, moves it on the board, and drafts the reply. They don't send autonomously. I still want to see how they get back to people.

Then something I did not expect happened. Founders started coming back happier. Not happy that I passed, but genuinely appreciative, because normally I don't have time to do proper competitive research on something I already know I'm going to pass on. Some of them replied saying it was a pity I wasn't investing, because they hadn't had feedback that good from anyone so far, and certainly not within three hours of sending the email.

I am now doing a better job of analysing and pre-filtering investment cases than I could do myself, or than I did when I had human colleagues.

The second was cockroaches.

Three days before flying to Cape Town, where we have a house, I found out we had cockroaches in the storage room. I asked Bianca, who is there, to find someone to deal with it before I arrived on the weekend. She said two days was not much notice, and maybe I'd just have to leave the house for a day the following week.

I thought, come on, this has to be solvable. So I complained to Thea about it, and told her she could go ahead and email companies and try to negotiate something.

Two hours later she came back with three companies that could come the next day, with the rates already negotiated. I just had to pick one.

What made it feel different was that it was not one email and a yes. It was a real back and forth: what time, what kind of cockroach is it, can you send a picture. For some of it she had to come back to me or to Bianca. But she handled the process. She watched for new replies, drafted new ones, asked me when she genuinely needed a decision, and sent on her own when she already knew what my answer would be.

The third was a negotiation, and this is the one that changed my mind about what these things are for.

I had been running DFINITY servers for about five years and decided this year it was time to get rid of them. I was working with a company that looks after nodes for providers like me for a service fee, and they told me they had done this before and could get me the best price in the market. Fine, I said, get me a quote.

These were originally a Tomahawk investment, and we had paid about a million dollars for them in 2020. Five and a half years later the quote came back: $190,000. But they couldn't sell yet, because my contract ran until June. Come back in a month.

So I came back a month later and asked for an updated quote. Two days later they came back with exactly $190,000 again. That made me suspicious, so I asked how much margin they were taking on top. They said they couldn't really say.

At that point I gave it to the agents.

By then I had upgraded the system, so they were a bit more thoughtful about researching a space and building relationships. What they worked out is that there are a lot of server brokers in Europe, but at the back end almost everyone sells to one of three actual buyers, the ones who store the servers and take them apart.

Within two hours I had an offer of $250,000, already a third above what the human company had given me.

After three days they had negotiated it to $410,000.

Why this mattered more than the money

The money was good. But the reason that third one stayed with me is that negotiating is a sales job. My agents built a relationship, researched the space, held a real conversation, and came back with something far more valuable than what I would have gotten myself, or what an established human intermediary had offered me.

That was the moment I stopped thinking of these things as a way to sort my email, and started wondering what else they could run.

Which is where the next post picks up.

This is part 1 of 3. Part 2: One hundred companies, three people. Part 3: Why I'm building Airmond.