Building agents that do more than chat
Why a durable, versioned agent with its own data and connectors is a different thing from a chat prompt — and how Amistio approaches it.
Most people meet AI agents as a chat box. You type, it answers. That is genuinely useful, but it stops at the edge of the conversation: close the tab and the work is gone. An agent that can actually help with recurring work needs three things a chat box does not have — a durable definition you can trust, a place to keep its own data, and a safe way to reach the tools you already use.
A definition you can trust
In Amistio, an agent is a workflow you assemble from typed blocks, not a free-form prompt. When you are happy with it you freeze it into an immutable, content-addressed version. That version cannot silently change under you: the same inputs run the same steps, and consequential steps can pause at an approval gate before anything irreversible happens.
Its own memory
A chat transcript is not a database. An agent that tracks shared expenses, keeps a watchlist, or records approvals needs storage that belongs to the agent and survives across runs. Amistio gives each agent a bounded, private data store, so the running record lives with the agent rather than scrolling out of a chat history.
Safe reach into your tools
The point of an agent is to act. Amistio connects to services like Slack, GitHub, Notion, and others through connectors you authorize with your own token, stored encrypted on your deployment. A token is only ever sent to the service it was created for, and a connector call goes through an egress guard. The aim is boring, predictable safety rather than surprising power.
Where this is today
The builder, the versioning, the agent data store, and the bring-your-own-token connectors described here run on the self-hosted embedded runtime today. A hosted multi-tenant service and cross-instance sync are planned, not yet available — we try to say clearly what is shipped versus what is on the roadmap, on every page.