Récif: What Happens After You Deploy an Agent

Deploying an AI agent is easy. Everybody talks about it.

What happens after gets a lot less attention. Monitoring tools exist (LangSmith, Langfuse, Arize). But monitoring is not the same as acting on the results: splitting traffic between versions, promoting automatically when quality improves, rolling back when it drops.

For ML models, the tools and practices exist. For agents, evaluation tools exist (MLflow Evaluation, Vertex AI Evals, RAGAS) with varying levels of depth, and platforms like LangSmith offer monitoring and CI/CD integration. But no open-source K8s-native solution closes the full loop: evaluate on live traffic, promote or rollback automatically, feed results back into the next version.

So we built Récif. Open-source, Apache 2.0, K8s-native.

The architecture has two layers. Corail is the agent runtime. Works standalone, just docker run. Récif adds governance on top: evaluation, canary, compliance. Think Docker vs Kubernetes. Start simple, add governance when you need it.

The core is champion/challenger deployment (shown in the video):

You push a new agent version. It becomes the Challenger. Istio service mesh splits traffic: 10% Challenger, 90% Champion. Flagger handles progressive delivery. MLflow LLM judges score every interaction on safety, relevance, tool accuracy, RAG groundedness. Challenger beats the threshold? Promoted. 100% traffic. Zero downtime. If not, automatic rollback. Users notice nothing.

User feedback is configured automatically per channel (Discord reactions, Slack replies, Google Chat, REST endpoint). These signals merge with MLflow observability traces. Every negative rating becomes a test case the next version must pass. The bar rises with every cycle.

Each agent runs as a single pod with its channel, tools, guards, memory, and knowledge base. Agents go where your users are: Slack, Discord, Google Chat, REST API, WebSocket, A2A. MCP servers and LangChain/CrewAI tools register natively.

Two adoption paths: marketplace agents for teams that don't code, GitHub scaffold + CI/CD for developers building custom agents.

Security: guard pipeline on every LLM call (prompt injection, PII masking, secret detection). Tools declare risk levels, dangerous ops require human confirmation.

Knowledge: Maree ingests from Google Drive, Jira, Confluence, S3. Hybrid BM25 + semantic search. The agent decides when to query, not a forced pipeline.

Every release is an immutable GitOps artifact (SHA-256, rollback to any version, ArgoCD-ready). Governance built in: cost ceilings, safety thresholds, real-time scorecards.

Multi-team (one K8s namespace per team). Multi-model (OpenAI, Anthropic, Gemini Enterprise, Bedrock, Ollama).

We're early. Building in public. But the vision is clear: treat AI agents with the same operational discipline we apply to any production service. And because every agent runs as a microservice behind a service mesh, the architecture is ready for multi-agent orchestration when it proves itself at scale.