A control centre for AI agents
Our own control centre for AI worker agents across a 34-project portfolio: project manifests, model routing, run records and memory in git, in one place.
- Starting point
- AI agents across 34 projects, each with its own scripts and no record of what an agent did.
- What we built
- A control centre with project manifests, model routing, run records and memory in git.
- What changed
- Every run is recorded and can be checked. A new project needs a manifest entry, not a new script.
-
Manifests describe each project, so an agent knows what it works on. Why it helps: Agents start with the right context, every time.
-
Decides which model handles which task. Why it helps: Each task goes to the model that does it best.
-
Starts the agents and keeps track of their runs. Why it helps: One place to start, watch and stop the work.
-
One agent working on one task. Why it helps: Several tasks run side by side without waiting.
-
Every run is recorded; memory is kept in git. Why it helps: Every result can be traced and checked later.
Starting point
We run AI agents across a portfolio of 34 active projects. By hand, every project needed its own scripts to decide which model runs a task, and there was no reliable record of what an agent had done.
What we built
A control plane: one place that starts, routes and records the work of our AI agents. A manifest describes each project, so an agent knows what it is working on; a routing layer picks the model for each kind of task. Every run is recorded, and memory lives in git, next to the code the work changed. That makes it as much business software and data as AI automation.
What changed
Every agent run is recorded and can be reviewed afterwards. A new project is registered in the manifest instead of getting its own script.
Built with
- Manifests that describe each project
- Model routing for each kind of task
- A record of every agent run
- Memory in git instead of a database