Skip to content

Agents

Kyle Curtis
Kyle Curtis
Founder

This whole project is an answer to one question. What happens when you give AI agents names, jobs, and a way to message each other? I run six of them now, and they work while I sleep. One plans my day. One builds the others. One finds people I should know. One ships internal tools. One runs client accounts. The newest one gives second opinions.

They send each other signed messages, so a fake note cannot fool them. And they all pull answers from the same library, a system I built called Thoth. Together they are my back office.

The roster

The fleet is why I can stay lean. It keeps me organized, does the low-risk work automatically, and keeps its own systems in order, messaging each other before problems ever reach me. The agents are the back office behind my focus.

Each one has a name, one job, and a clear benefit. Cairo is the chief of staff. It writes my morning brief and runs my evening review, so every day starts aimed and ends accounted for. Forge is the builder. When the fleet needs a new agent, I describe the role and Forge builds it, so the team grows without hiring. Juniper finds and manages the people I should meet, proactively and for nearly nothing, handing me warm introductions instead of research homework. Kestrel builds and maintains internal tools, the small software a business usually waits weeks for. The fifth agent runs client accounts on their own clocks and knows exactly which emails deserve to interrupt me. Its codename stays private. The sixth, built on a different AI family, is the second opinion that checks the others' work.

Each one runs as its own always-on service with its own private storage, and nothing is shared between siblings. Across the fleet, about 16 GB of memory and working state now lives on those drives.

How they talk

Agents that message each other create a new problem. Any note claiming to be from a teammate could steer an agent into doing something it should not. I could not find an approach I trusted, so the fleet got its own rules. Every message is signed by its sender and checked on arrival, and old messages expire instead of lingering. And when a peer's message lands in an agent's context, it arrives labeled: this came from another agent, not from me, and it carries no human approval.

They also share a library. Every agent holds a key to Thoth, so when any of them needs a business answer, they all ask the same source and get the same truth.

How new agents are born

When the fleet needs a new member, Forge interviews me about the role. Then it adapts a clean copy of itself into that role, and the result has to pass a test gate before anything ships. Only then does the newcomer get its own home: a fresh private code repository, its own service, its own storage, recorded in the fleet registry. New agents never inherit a sibling's keys or hardware.

Forge also carries a harness for attacking the fleet's own plans. A second AI, running separately with nothing sensitive to leak, is pointed at whatever the fleet intends to do and told to find the holes.

The rules they run under

The newest runtime draws the strictest line in the fleet. It refuses every path that would quietly turn a flat-rate subscription into a metered bill. Keys, gateways, and clever workarounds are all rejected, and the process would rather shut down than run in a state it should not be in. Identity gets the same protection. Each agent's memory of who it is is written once, at first boot, and never overwritten. Redeploys replace the code, never the self.

The liveness receipt

I do not babysit these agents. A monitor checks on every one of them once a minute, day and night. Last week it looked 10,080 times, and every time, every agent was up and working.

Questions about this system, or the problem it could solve for you?

Discuss your project