Three YC-backed founders have launched Cekura, a testing and monitoring platform for voice and chat AI agents. The YC F24 startup addresses a critical challenge in AI development: ensuring conversational agents behave correctly when prompts change, models are swapped, or new tools are added. Traditional QA methods like manual spot-checking or scripted tests fail to capture the complexity of real user interactions.

Cekura's approach uses synthetic users to simulate conversations and LLM-based judges to evaluate responses across entire conversational arcs, not just individual turns. The platform offers three key capabilities: scenario generation that bootstraps test suites from agent descriptions, mock tool platforms that simulate API calls without touching production systems, and deterministic test cases that eliminate the stochastic nature of LLM responses. This ensures failures represent real regressions rather than random noise.

The startup positions itself against tracing platforms like Langfuse and LangSmith, arguing that conversational agents require session-level evaluation rather than turn-by-turn analysis. Cekura offers a 7-day free trial with paid plans starting at $30/month. The team invites the developer community to share their experiences with testing behavioral regressions in AI agents.