Architecture
System Overview
How UnderOcean's platform layers fit together.
Platform layers
UnderOcean is organized into a small number of clearly separated layers, each with a single responsibility:
- Orchestration Engine — executes agents and flows as governed, stateful workflows, with built-in checkpointing so multi-step work can pause, resume, and recover safely.
- Model Provider Layer — a unified gateway in front of every supported AI provider, so agents and flows are never locked into a single vendor and can switch models without any workflow changes.
- Knowledge & Retrieval Layer — a vector-based knowledge store, isolated per project, that powers Agentic-RAG — grounding agent answers in your own documents rather than relying on model memory alone.
- Secure Multi-Tenant Data Layer — every read and write is scoped to an organization and project boundary, enforced beneath the application layer so tenant data cannot leak across boundaries.
- Background Intelligence Layer — long-running work such as document ingestion and context summarization runs asynchronously so it never blocks a user-facing request.
- Real-time Streaming Layer — every execution streams progress, reasoning steps, and tokens live to the UI over a persistent connection.
- Capability Surface — one table of project capabilities (read, build, run) is offered to people through the AI Assistant and to external AI clients through an OAuth-secured MCP endpoint. Both go through the same permission checks, rate limits and audit trail as the rest of the platform.
- Versioning Layer — agents and flows keep a bounded history of snapshots and a revision counter, so concurrent edits from the UI, the assistant, MCP clients and the API cannot silently overwrite each other and any change can be undone.
Agents and flows share one engine
An agent is the simplest possible workflow — a single step. A flow is a larger, user-composed workflow built from many steps. Both run on the same orchestration engine, which means every platform capability — streaming, checkpointing, visibility into reasoning — works identically whether you're running a single agent or an entire multi-agent pipeline.