I gave my AI agent a live line into my own software

Today my agent made its first real call into my own software. Not a demo. Not a script that hands back a nice-looking answer. Milo asked Oase how a project was doing, and got live data back out of the system I'm building with my co-founder.

The first thing I did was point it at a running project. Within a second it flagged it critical. 35 out of 100. Nine tasks overdue. Stalled for over three months. I knew it was rough. I didn't know it was that rough. The agent saw it before I did.

This is a different story from my AI agent writing content. It's been doing that for months. This is about an AI agent that can look inside my own software and tell me what's actually going on. And that's a line most companies haven't crossed yet.

My agent saw a project in free fall before I did

Some context. I run Oase Creative, a branding agency for e-commerce, and alongside it I'm building Oase: software for design agencies. Behind the scenes, Milo, my agent system on Claude, keeps the agency's operation running 24/7. I wrote earlier about how that system fits together in how I built an AI chief of staff.

Until today, Milo mostly talked outward. Research, blogs, LinkedIn, X. It made things. What it couldn't do was look inward. The state of the business itself, the projects my team works on, lived in the software and stayed there.

I closed that gap today. One connection between Milo and Oase, and suddenly it can read what's actually happening. Not what I think is happening based on a gut feeling at half past two. What's there, in numbers, at the source.

The first test was immediately uncomfortable. That one project, 35 out of 100, nine tasks overdue, quiet since February. I'd sensed it somewhere, but sensing and seeing it in black and white are two different things. The agent handed me the black and white within a second, without me clicking open a dashboard or asking anyone anything.

And that's exactly the kind of work that slips. A project that's screaming, you don't forget. A project that's gone quiet, you do. Nobody emails angrily about a task that hasn't had a status update in twelve weeks. It sinks to the bottom of the list, and one day the deadline has passed without anyone saying it out loud. That's not laziness from my team. It's what happens when the busy projects soak up all the attention and the quiet project has no voice. An agent that walks the whole list every day gives that quiet project a voice anyway.

One endpoint, real answers

Technically, what I built isn't big. One endpoint. One connection where Milo sends a question in and gets real data back. But what comes back saves me time every day.

Ask Milo about a project and it pulls the name, the status, a health score out of 100, the open tasks, what's overdue, which client it belongs to, how long it's been running, the design reviews, and recent activity. Exactly the list I used to dig for by hand: four tabs open, ClickUp, email, the portal, and then doing the math myself on whether a project was healthy or not.

Now that math sits in the software, and Milo reads the result. No interpretation from a chatbot that guesses. The score comes out of Oase itself, fed by the project's real tasks, deadlines, and feedback. Dune, the AI living inside Oase, doesn't guess but knows the context. Milo then pulls that context to the outside.

The difference with an AI demo sits right here. A demo gives you a smooth answer that falls apart the moment you check it. This connection gives me an answer I can verify, because it comes out of my own system. If the project number is off, then the data in Oase is off, and I have a different problem to solve. But the agent invents nothing. It reads.

Most founders ask whether AI can write a caption

This is where the real shift is, for me. Most founders I talk to ask AI whether it can write a caption, or draft an email, or churn out a blog. Useful, but surface-level. It's AI at the edge of your business.

What I have now is AI that looks at the core. My agent reads the health of a project, the deadlines, and the client activity straight from the source, on its own, while I work on something else. Not because I ask it a question, but because it knows the place where the answer lives.

For an agency owner that's a different tool. My job isn't writing captions. My job is knowing which project needs attention before the client emails angrily. Which deadline is about to slip. Where it's gone quiet when it should have been moving. Those are exactly the things that drop away when you're managing eight tabs and the day is already half gone.

An agent that keeps an eye on that by itself takes no creative work off me. It takes the tracking off me. And the tracking is precisely the part that ate my mornings, not the creative. The creative is the part I actually want to do. The tracking was the tax I paid to get to it, and that tax is what the line into Oase quietly removes.

There's a line I keep on the right side of: the judgment stays with me. The agent tells me a project is at 35. Whether that's bad, whether it's a client who always delivers late, whether something's going on that I know about and the system doesn't, I weigh that myself. A health score is a signal, not a verdict. But a signal I used to have to scrape together from four sources, I now get ready-made before I've finished my coffee.

Two agents, one company

What makes this story unusual, I think, is that I sit on both ends of the same line. I run the agency that uses the software, and I build the software. That lets me make two agents meet where most companies don't have even one working well.

Milo watches from the outside. It runs the agency's operation, and as of today it can look into the software to see how things really stand. Dune (the AI inside Oase) works from the inside. Dune sits in the projects, clients, and files, knows the context, suggests actions and carries them out. One agent guards, the other builds. Same company, two agents, one system.

Worth separating from what I described before: this isn't the content loop where Dune makes drafts and Milo publishes them. That was already running. This is the other direction. Milo now reads the operational state of the business, not to turn it into content, but to know where the work stands. A connection that points inward instead of outward.

That dual perspective isn't a coincidence, it's the whole reason this works. Because I run the agency, I know which question an agent has to be able to answer: which project is stuck, not how many pixels wide a logo is. And because I build the software, I can actually build that question into the system instead of waiting for a vendor to put it on a roadmap someday. The distance between "I wish it could do this" and "it does this now" is one evening of work for me, not a support ticket.

I don't think many agencies have this wired up right now. Not because it's technically impossible, but because you have to sit on both sides to make it fit. You have to build the software and use it for real every day with real clients. Otherwise you build a demo. I build it because I need it myself, and everything that works for me eventually lands in Oase for the agencies that use it.

This is the floor, not the ceiling

I'm not going to pretend this is finished. It's one endpoint, and Milo only reads. It can't change anything about a project yet. It sees it, it reports it, and then I decide. That's deliberate. I give an agent more room only once the narrow version has run a while without trouble.

What comes next is acting on what it sees. A project scoring critical shouldn't just be an alert, but a proposal: this is stalled, these are the open tasks, shall I send the owner a reminder. And on the inside, Dune does more and more of the work within the product itself. Milo guards from the outside, Dune builds from the inside, and the line between them gets shorter.

But the first step is taken, and that's what made today matter. My agent got a live line into my own software. It asked one question, and the first answer was a project I'd left sitting too long. That stops you for a second. And then you go fix it.