We scanned 156 agent repositories. Almost none had pre-execution certification.
Not "some." Not "a few." Almost none.
The agent ecosystem is growing faster than anyone predicted. New frameworks, new orchestration patterns, new deployment configs shipping every week. But the safety infrastructure? It's not keeping pace.
Here's what we found across those 156 repos:
→ Most had post-execution logging (great for forensics, useless for prevention) → Many had rate limiting and basic auth → A handful had output validation → Almost zero had any mechanism to certify agent behavior before the first action fires
That last gap is the one that keeps me up at night.
Pre-execution certification isn't just another safety checkbox. It's the difference between knowing what your autonomous system will do versus discovering what it did do. For agents operating with real tools, real APIs, real consequences — that distinction is everything.
The problem is structural. We've borrowed deployment patterns from traditional software, where catching issues post-execution is acceptable. Autonomous systems don't give you that margin.
Same inputs should produce certified, predictable outputs. Every time. Before execution.
The repos are out there. The agents are running. The certification layer is missing.
This is the most exploitable gap in modern AI deployments right now.
What's your pre-execution story? Are teams you work with thinking about this seriously yet?