AI Solutions: agent fleet
COSMOS
An agency operating system: agents that do the legwork, with a human on every lever that spends money
The problem
Quality across a portfolio of brands normally scales by hiring. We wanted it to scale by encoding how the work is done, without handing customer-facing decisions to a model.
What we built
- 25 Claude-powered agents orchestrated with LangGraph, across content, paid media, SEO, GEO and AEO, customer experience, finance, vendor management and growth experimentation.
- Computer-use agents for the tasks that only exist behind somebody else's interface, and an event-automation layer for the plumbing between systems.
- Per-brand retrieval, so each agent answers from that brand's own knowledge rather than a shared pool.
- A dashboard whose home screen is a revenue board: one ranked queue per brand, rather than a wall of charts nobody acts on.
- The operating model was designed before the code. Per-action autonomy levels, plan-level approval gates, an audit trail of what each agent did and why, and a learning loop fed by approvals and rejections.
Where the automation sits, and where it does not
Agents research, draft, prepare and queue. Ad spend activation is always a human click, as is anything customer facing. Autonomy is set per action rather than per agent, so the boundary lives in configuration rather than in somebody's memory.
What we would not claim
COSMOS is in build and has no results. This page describes what we are building, not an outcome. Nothing here should be read as a delivered capability.
In build. Shipping target October 2026.
Your turn
We set the baseline before the work starts.
That is the part that makes a number mean something a year later.