Introduction

What is Récif?#

Récif is an open-source agentic platform that separates how you build agents from how you deploy and manage them. It provides enterprise-grade governance, evaluation, and operations for AI agents — regardless of the framework or cloud provider you use.

The platform is built on two independent layers:

• Corail (the coral) — an autonomous Python agent runtime. Each agent runs in its own container with its own LLM, tools, memory, and channels. Corail works standalone — no platform required.
• Récif (the reef) — the governance and operations layer. A Go API + Next.js dashboard that adds evaluation, releases, governance, monitoring, and multi-tenancy on top of your agents.

Two Access Paths — At the Same Time#

This is the key insight: Corail agents are autonomous. They don't need Récif to run. But they can also be governed by Récif. Both happen simultaneously on the same agent — this is not a choice between one or the other.

Direct Access — Developer / Integration#

Your application talks directly to the Corail agent. No platform in between. This is ideal for:

• Developers building an AI feature into their app
• API integrations — your backend calls the agent's REST API
• Slack/Discord bots — the agent connects its own channel
• CI/CD pipelines — automated agents triggered by events

Your App → Corail Agent (port 8000) → LLM → Response

Each agent exposes its own REST API, SSE streaming, and channel connectors. It handles its own tools, memory, and knowledge bases. Zero dependency on the platform.

Platform Access — Enterprise / Mass Adoption#

At the same time, users can interact through the Récif dashboard or API. The platform proxies requests to the same agents and adds governance, monitoring, and evaluation. This is ideal for:

• Enterprise adoption — give every employee access to AI agents through a single portal
• Multi-team environments — each team has its own namespace, agents, and permissions
• Production governance — eval-gated releases, canary deployments, quality gates
• Compliance — audit trails, cost tracking, guardrail policies

User → Récif Dashboard/API (port 8080) → proxy → Corail Agent → LLM → Response
              │
              └── Governance, Eval, Monitoring, Feedback, Releases

The platform adds:

• Authentication & RBAC — who can access which agents
• Evaluation pipeline — 14 MLflow scorers gate every release
• Canary deployments — 10% → 50% → 100% with quality gates
• Feedback loop — user thumbs up/down feeds back into evaluation datasets
• AI Radar — fleet-wide health monitoring and alerts
• GitOps releases — every config change is a versioned, auditable artifact

Both Paths, One Agent#

This is not a choice — both paths coexist on the same agent, at the same time. A CI/CD pipeline calls the agent directly via REST while an employee chats with it through the dashboard. The agent doesn't know or care which path the request came from. Récif layers governance on top without touching the agent code.

Architecture at a Glance#

Each agent runs as a Kubernetes Pod with its own Corail runtime. Récif manages them through a kubebuilder operator that reconciles Agent CRDs into Deployments, Services, and ConfigMaps.

Core Capabilities#

Next Steps#

• Quickstart — install Récif and deploy your first agent in 10 minutes
• Architecture — deep dive into the two-layer architecture
• LLM Providers — configure OpenAI, Ollama, Vertex AI, and more
• Evaluation Guide — the eval-driven lifecycle explained