We hit a wall testing AI agents, agents simulations works better

We've been working with teams building AI agents (agentic systems, with actual execution) But here's the thing: everyone says “agents are the future,” yet no one really knows how to test them. Some teams are manually walking through conversations, others are just shipping and "vibe checking" what comes back. Both break down at scale. The real problem? We’re testing agents like software, but agents don’t behave like software. They make decisions, adapt, escalate, reason across contexts. They're more like processes than functions. Rogerio, our CTO, wrote up a deeper dive on how we see the future of agent testing, and why agent simulations (not hardcoded flows) are becoming the new unit tests for AI systems. We built LangWatch scenario to let teams simulate real-world agent behavior and catch regressions early on. Would love feedback from folks who’ve been burned by this or hacked together their own simulation setups.