An agent is a DB row; the runtime is an ephemeral k8s Job that exists only while a task runs. Scale by traffic, not by agent count.
These are the living development docs for the greenfield re-platform. One idea runs through everything:
The agent is always available — a row in the DB. Its runtime exists only while a task runs.
Today's 1.x is an AWS-locked, request-centric RAG chatbot. 2.0 is rebuilt to become the AI agent for CMS / e-commerce stores — OpenCart first — cheap enough to give every store its own.
New here? Start with the Overview, then the Layered slices.
Three moving parts: api runs the agent's brain per turn, worker is an on-demand worker for its hands (bash / fs / browser), and LightRAG serves knowledge — all on one shared Hetzner k3s cluster, backed by Postgres and Redis.
Go deeper in the Architecture and the Runtime model.
What ships, version by version (full detail + owners on the Plan page):
| Version | What gets added |
|---|---|
| v0.1 | Scaffolding — repo, CleanSlice skeleton, infra/setup |
| v0.2 | Core (no-runtime) — agent, llm, chat, memory, orchestrator |
| v0.3 | Ephemeral Light runtime (MVP) — tasks, worker, events, file |
| v0.4 | Browser runtime + billing/usage |
| v0.5 | Heavy + Warm + scheduled tasks |
| v0.6 | Knowledge — LightRAG service |
| v0.7 | Frontend — app + admin + SDK |
| v1.0 | Production — IaC, observability, hardening |