What happened
A recent InfoQ webinar brought together several specialists to discuss what can go wrong when agentic systems are put into production. The discussion stressed that these systems can produce convincing errors without failing outright, which makes problems harder to spot than in traditional software.
The risks mentioned included semantic drift in LLM pipelines, embedding drift in financial compliance systems, accidental retrieval of sensitive documents in RAG flows, and exposure of API secrets. It was also noted that malicious repository configurations can lead to unauthorised code execution.
The speakers argued for automated code review, continuous model evaluation, mutation testing, and stronger observability. Their recommendation was to treat AI agents like new team members, with their own identities, limited access, and controlled dependencies, while the final decision to deploy remains strategic.
Key points
- Agentic systems can fail silently without stopping the service.
- Semantic and embedding drift can create hard-to-detect errors.
- Manual code review no longer scales with the volume produced by agents.
- Model evaluation should be continuous and built into operations.
- Security and privacy need stronger controls over agents and secrets.
Why it matters for your organisation
For companies and public organisations, the message is that adopting agentic AI cannot rely only on initial testing or trust in model behaviour. Engineering, security and governance need to work together to reduce operational risk, protect sensitive data and prevent subtle errors from affecting business decisions.