REC.Event Recap: From Enterprise RAG to Agentic Workflows to OpenClaw Safety

On August 29, CADSEA welcomed Dr. Rujun Gao, Co-Founder and CEO of Encando.AI, for an online talk on the engineering decisions behind enterprise AI systems. The conversation followed a practical progression: building reliable RAG, designing agentic workflows, and managing the risks that come with greater autonomy.

From a RAG Demo to a Production System

A RAG demo can retrieve a passage and generate an answer. An enterprise system must do considerably more. Dr. Gao discussed the challenges of parsing complex documents, keeping indexed content current, and ensuring that users can retrieve only information they are authorized to access.

She described a layered retrieval approach that combines semantic and keyword search, followed by re-ranking to improve relevance. Evaluation with human-verified questions and end-to-end observability help teams understand where an incorrect answer originated. Citations make an answer traceable, but they cannot make outdated or incorrectly retrieved information accurate. Throughout these decisions, teams must balance answer quality, latency, and cost.

Designing Agentic Workflows

As AI moves from answering questions to completing multi-step tasks, the design focus expands to the agent’s entire working environment. An agent may need to decide when to retrieve more information, which tool to call, how to use memory, and when to retry or escalate to a person.

Dr. Gao introduced patterns including ReAct, Planner-Executor, and supervisor-based multi-agent workflows. She also discussed how protocols such as MCP and A2A can help connect agents with tools and other systems. The appropriate level of autonomy depends on the task and the controls around it.

Building Safety into the System

Persistent memory, third-party tools, and direct execution make agents more capable, while creating new opportunities for prompt injection, untrusted code, and information leakage. Dr. Gao outlined safeguards across three areas: identity and authentication, policy and prompt defenses, and isolated execution with limited credentials. She also emphasized the need to govern what an agent remembers, where that information came from, and when it should be updated or deleted.

Thank you to Dr. Gao for sharing these practical lessons, and to everyone who joined the discussion. The session offered a useful question for any team building with AI: as systems gain the ability to act, how do we keep their information, permissions, and actions reliable?

Watch the ⁠full recording on YouTube.