Blue Guardrails GmbH

Blue Guardrails is a Berlin-based software company developing a reliability layer for AI agents in production. The company helps product and engineering teams monitor, evaluate, steer and continuously improve agent behaviour. Blue Guardrails was founded with the aim of enabling teams to use AI responsibly and to address core technical challenges such as hallucinations, robustness, human oversight and fairness.

The platform analyses production traces and detects issues in agent outputs and action sequences, such as hallucinations, faulty tool calls, or deviations from the system prompt. In addition to quality, it also monitors cost, latency and tool usage; individual traces can be examined in detail with labels and explanations, and the development of error rates can be tracked over time.

For evaluation purposes, production traces can be replayed to compare different models, reasoning settings, prompts and tools in terms of error rate, latency, token consumption and cost, without altering the live agent. With active prevention, Blue Guardrails intervenes during an ongoing run: detected issues are fed back to the agent as contextual feedback, allowing a response to be corrected or the next step to be steered before users or downstream applications are affected. In addition, an optimisation agent analyses recorded production conversations to investigate the causes of errors, tests possible corrections, and provides evidence-based recommendations on prompts, models, tools and agent setup.

The offering is aimed at engineering and product teams operating agent and RAG applications. Custom error categories and domain context can be defined so that detection matches the specific use case and focuses on the relevant error patterns. The solution is developed in Berlin and is available with EU data residency, as a private cloud variant, and as an on-premises deployment.

The company is led by co-founders Miriam Kümmel and Mathis Lucka. Miriam Kümmel brings experience with large language models and production AI systems from areas including legal publishing, news media, finance and public administration, and previously led forward-deployed engineering teams at deepset.