AI in Production: Experts Flag Silent Failures, Harness Engineering, Agent Governance

In an InfoQ webinar, senior engineers from ThoughtWorks, privacy and financial AI specialists dissected operational failures of agentic systems: semantic and embedding drift, data leaks via prompts and RAG, and code review bottlenecks from AI coding assistants. Recommendations include automating checks in CI/CD, versioning MCP dependencies, and giving agents distinct identities with tiered access.

• Semantic drift in text-to-SQL agents silently corrupts downstream business analytics
• A background embedding model update spiked false positives while health checks stayed green
• Manual line-by-line code review cannot scale with AI-generated production code
• Agents should get non-human identities and isolated API secrets in sandboxes