Reference
Architecture map
Optimal shape for this product: harness-driven agent + durable orchestrator + React for Instagram connect and run monitoring.
flowchart TB
clients[Clients: Harness MCP + React UI]
interface[Interface: MCP + REST + SSE]
orch[Orchestrator: state · policy · runs]
agent[Agent loop]
model[Model adapter]
tools[Skills / tools]
policy[Policy gate]
world[World: IG · media · files]
memory[Memory + evals]
clients --> interface --> orch
orch --> agent
agent --> model
agent --> tools
tools --> policy --> world
orch --> memory
orch -->|events| clients
Layer ownership
| Layer | Owns | Must never |
|---|---|---|
| Clients | Human steering & display | Own business rules or secrets |
| Interface | Stable contracts for any client | Bypass policy for convenience |
| Orchestrator | Truth, transitions, run records | Pretend to be the LLM |
| Agent loop | Plan / act / observe within budgets | Call Instagram without a gate |
| Model adapter | Vendor-agnostic completions | Leak into every skill file |
| Skills | Structured capabilities | Own DB schema or OAuth |
| Policy | Allow / deny side effects | Trust model “intent” alone |
Dual client surface
MCP (any harness)
Cursor, Claude Code, scripts — interact with skills and inspect runs.
React (product ops)
Connect Instagram, approve drafts, monitor live agent activity.
Repo mapping
Today: Express + SQLite + LangChain skills + MCP skills server are the right instincts. Gaps: OpenAI-compatible model adapter, run events, wired UI, real OAuth/media/publish behind policy.