Agents start as contracts: role, inputs, outputs, schedule, state, observability and allowed side effects.
Technical core / SAI layer
SAI Agent Factory
SAI is the portable agent factory behind the visible prototypes. It turns structured intent into missions, blueprints, validated execution, context writes, observation loops and gated outputs. Robots and avatars are surfaces. SAI is the system layer underneath.
Persistent memory belongs in structured context, not in a single chat. Observations become reusable evidence.
SAI routes tasks to tools, documents, cameras, scouts, web interfaces and future physical devices.
Public messages, notifications and actions can pass through policy, review and human-facing constraints.
Each run should leave a trace: what was requested, what was checked, what changed and what remains open.
Voice, STT, TTS, RAG, vision, scheduling, maps and device bridges become shared organs for many agents.
Visible runtime proof
The older Snapy OS page is the closest public artifact for this layer: a visible AI agent runtime with skill orchestration, event stream, web and Telegram surfaces, tool calls and a path toward physical device presence. It is separate from the robot characters, but it is the system idea that later powers them.
Factory flow
1. Intent
2. Requirements
3. Blueprint manifest
4. Tool selection
5. Validation
6. Schedule or run
7. Context write
8. Trace and learning loop