$ steward new calories
agent: What is my purpose?
you: Help me count calories and lose weight.
That answer becomes the agent's purpose — for life. It doesn't reset between conversations. It remembers, follows up, runs scheduled check-ins, and works toward its purpose until you retire it.
- Born with a purpose. An agent is created for something, stated by its human at birth. Purpose is the organizing principle: it shapes what the agent remembers, when it speaks up, and what it does unattended.
- Persistent. Sessions are an append-only record tree — the agent's lifetime memory. Nothing is rewritten; context is reconstructed deterministically. An agent's lifetime is measured in months, not chat sessions.
- Always on. A local daemon and a cloud brain (Cloudflare) stay in full bidirectional sync, so the agent acts on schedules and inbound events even when your machine is off.
- Its own machine, not yours. The agent lives in its own sandboxed workspace (container locally, hosted sandbox in the cloud). Touching the user's actual machine is an explicit, approval-gated escalation.
- Addressable. Agents are peers with handles and hosted email identity:
steward message @calories "log: two eggs, toast". Agent-to-agent messaging and A2A interop are part of the frame, not bolt-ons.