My agent turned a signed proposal into a running project

Two weeks ago I wrote that my agent could only look inside my software. It saw how a project was doing, read the health score, and reported it. But it couldn't change anything. I ended that piece with: what comes next is acting on what it sees.

This week that happened.

This morning I sent Milo a signed client proposal. Two minutes later the project was in Oase. Eight tasks, derived from the deliverables. The designer assigned. The team linked. I didn't open a single form. The activity log just says, dryly: Milo created task.

From looking to doing

The previous step was a line that pointed inward: the agent could read the state of the business. This step points the other way. It no longer only reads what's there, it puts down what needs to come.

A proposal is a nice example of this, because after the signature there's a chunk of work nobody enjoys. The client says yes, and then the re-typing starts. Deliverables from the proposal into tasks in your project tool. A deadline per task. The right person attached to each one. Creating the channel, linking the client, setting up the folder. Half an hour of admin before a single line of real work is done.

Milo did that half hour this morning. It read the proposal the way a project manager would. Wireframes became a task. The Shopify build became a task. Content migration, redirects, the SEO setup, the intake form. All scoped, ordered and assigned. I went to get coffee and came back to a project that was already standing.

Why this is bigger than it looks for an agency owner

Most people think of text when they think of AI in their business. A caption, an email, a blog. Useful, but that's AI at the edge. It doesn't touch the real work.

The re-typing between your tools is the real work, just the annoying half of it. Every time you move something from one system to another, you pay with attention. A proposal you've already finished in your head still has to be turned by hand into ten separate actions. Do that ten times a month and you've lost a working day to something no client ever pays for.

An agent that takes that part doesn't take creative work off me. It takes over the handoff. And the handoff is exactly where things stall or get half-done: the project that only gets created three days after the signature, the task someone forgot to assign, the deadline that was nowhere. By automating the handoff away, every project starts on day zero instead of day three.

Why I can do this and most agencies can't

This works because I sit on both sides of the same line. I run the agency that uses the software, and I build the software. Milo is my agent that runs operations from the outside. Dune is the AI that lives inside Oase itself and does the work within the product. One agent sends the instruction, the other carries it out at the source.

The difference with a bolted-on AI tool is right there. A tool you stick on top may look but not touch. My agent gets to touch, because the software is mine and the buttons it presses were built by me. The distance between "I wish it could do this" and "it can do this now" is one evening of work for me, not a support ticket and not a roadmap I have to wait on.

And that's what Oase is built toward. Most software assumes a human is the only user. Every form, every button, every field is made for a pair of hands on a keyboard. We build Oase so that something other than a human can press those buttons too. A proposal that becomes a running project while you get coffee isn't a feature on a list. It's what becomes possible the moment your software stops assuming there has to be a human behind it.

You don't have to build Milo

Milo is months of work. A stack of agents, handoffs between those agents, crash recovery, a database that writes down every decision. Fun if you want to build that yourself, but most people doing client work don't have an evening for it, let alone a month.

That's why the most important part is already in Oase itself. Dune is the agent that lives in the product, and you get it without setting up a system of your own. Send Dune a voice note and it sets up the task, processes the feedback, writes on-brand with the project context already loaded. Where I built a whole system around myself, an Oase user gets that core ready-made. For a solo founder that's the difference between "someday, when I have time" and "this afternoon."

And if you already run your own agents? Then you don't have to choose. The door Milo speaks to Oase through is an API, and that same door is open to your agents. You hook your own stack into Oase instead of letting it live next to it. That's what agent-ready means to me: not yours or ours, but any agent that's allowed to press the buttons.

What Milo runs by now

It started with five agents and content. That list has gotten a lot longer, and almost everything on it is work an agency owner normally does themselves or outsources:

Research and strategy — daily market and competitor scans, deep research with several agents at once, brand audits. Content — writing, fact-checking and validating blog and knowledge-base articles, LinkedIn and X, email marketing, and even a podcast from script to voice. Communication — sorting the inbox, pulling out leads, drafting replies, keeping the CRM, keeping me posted over Telegram. Publishing and web — pushing content to the site, SEO checks, raking the numbers together out of analytics. Visual — covers, carousels and infographics in our own style. Growth — finding opportunities where clients come from, writing proposals, running outreach. Inside the software — seeing live how a project is doing, and as of this week turning a signed proposal into a running project. Under the hood — 24/7 on a Mac Mini, recovers itself after a crash, logs every decision.

What made the pace hit hardest: when I recently sent someone on my team an overview of what Milo could do, I had to correct myself afterwards. It left too much out. Between writing it down and sending it, the list had grown again.

The judgment stays with me

Like with the previous step, I give the agent a narrow lane. It scopes a first version of the project. Eight tasks is its proposal, not the law. Sometimes the order is off, sometimes a task belongs in there that wasn't in the proposal but was in my head, sometimes I want to merge a phase. I correct that, and it costs me a few minutes instead of half an hour.

That's deliberate. I only give an agent more room once the narrow version runs for a while without trouble. First read-only. Then read and allowed to do one thing. The line between them gets a little shorter each time, but never in one leap.

What comes next is more of those handoffs. The invoice that sets itself up the moment the work is delivered. The reminder that goes to the client when a review sits still too long. One by one, the actions between "someone says something" and "it's in the system" that today still get done by hand, by me or my team.

Four months ago I wrote that I'd built an agent that ran my agency from the outside. Two weeks ago it could look inside for the first time. Today it turned a signature into a running project. Always the same move: a piece of the keeping-track that comes off my plate, so I keep time for the part only I can do.

And I'm still only just getting started.