Work / AI infrastructure
Wrasse
A personal AI agent running inside WhatsApp and Telegram, operated for the user rather than self-hosted.
State
In build
Domain
AI infrastructure
Our role
Built & run
Still ours
On call
The problem
Agentic systems are easy to prototype and hard to put in front of a person. The moment an agent can act on someone's behalf, the question stops being whether it is clever and becomes whether it can be trusted with the account it is holding.
What we built
A personal agent that lives in the messaging apps people already use, operated as a service rather than self-hosted. Every action it takes is checked against the owner's rules before it executes, and written to a replayable record.
The hard part
The permission model, not the reasoning. An agent that asks permission for everything is useless; one that asks for nothing is unshippable. We first evaluated the rules after the fact, for audit — and had to move the check in front of execution, which meant every tool the agent could reach had to declare what it would do before it did it.
Where it runs now
In build, running against real accounts under supervision.
What this means for your project
It is the evidence behind the AI feature and copilot retrofit, the knowledge assistant, and the AI-readiness audit. If you are putting an agent in front of a customer, this is the part we have already solved.
The permission model is the product.