Our own build · rebuilding

Mission Control

The dashboard we run our own operation on — a task queue, a set of AI agents with their own memory, and every capability they have exposed as a logged tool. It’s being rebuilt on a new engine; this is how it’s put together and where it’s going.

In short

What it is
An internal dashboard for running a fleet of AI agents — ours
Status
Being rebuilt on a new engine
Parts
Task queue · per-agent memory · CRM module · MCP tool layer · run log
Why it exists
We run our own intake and follow-up on it
Public
Not yet — this is the architecture, not a live demo

Why we built it

Running more than one agent by hand falls apart fast. You lose track of what’s queued, agents step on each other, and when one does the wrong thing there’s no record of why. Mission Control is the layer that makes a fleet of agents something you can actually operate: one queue, one log, one place to see what every agent is doing and stop it if you need to. The design principle is the same one we sell — the agent proposes, a human can always see and override.

How it’s put together

Task queue
Work comes in from a form, a webhook, a schedule or another agent, and lands in one queue with a priority and a state. Nothing runs on a cron nobody can see — every job is a row you can inspect, retry or kill.
Per-agent memory
Each agent has its own context store: the script it works, what it learned about a contact, the last few runs. Scoped so one agent can’t read another’s state unless we wire it to.
CRM module
A thin contact and deal store the agents read and write directly, plus a sync to the real CRM. Agents don’t get raw database access — they go through this.
MCP tool layer
Every capability an agent has — send an email, book a slot, look something up, update a record — is an MCP tool with a defined input, a permission, and a log line. Adding a capability means adding a tool, not editing an agent.
Run log
Every step of every run is recorded: the input, the tool calls, the model output, the decision. When something goes wrong we can replay it instead of guessing.

What the rebuild changes

The first version worked but was welded together — the agent logic, the queue and the tools all knew too much about each other, so every change was risky. The rebuild separates them: agents are configuration, tools are independent and individually permissioned, and the queue doesn’t care what runs on it. When it’s back we’ll publish the write-up with the parts that are safe to open-source.

Want a layer like this over your own agents? Tell us what you’re running.