Methodology for Designing AI Agents in Enterprise

Many teams fall into the "prototype trap": in a sandbox the AI agent looks flawless, but in production it breaks under load, network connectivity issues, drift, and hallucinations. Classic systems analysis falls short here: there are no interfaces anymore, and blind trust in an LLM leads to incidents.

The talk presents an original framework for designing production-grade AI agents, based on hands-on experience deploying agentic systems at the country's largest bank. We'll cover the analyst's shift from screens to deterministic skills, specifications, and safe execution environments.

You will learn:

• How to filter out non-agentifiable processes using a risk funnel and an autonomy matrix.

• Why pure ReAct is dangerous in production and how to build a hybrid graph (StateGraph + MCP + HITL).

• How to move from static tests to LLM-as-a-Judge and SDD.

• How to protect the environment: 4 layers of defense and model risk management.

You'll walk away with a checklist for designing reliable enterprise agents.