G2 defines trust layer framework for probabilistic AI agents

Alexis Zhang, Chief Product Officer at G2, defined the 'agent trust layer' as a framework for verifying probabilistic AI agents that operate in black boxes. The framework consists of three components: verification via metrics to ensure agents achieve desired outcomes, guardrails to block wrong actions, and observability to track deployed systems. Zhang noted that writing evaluations is difficult because it requires humans to verify the verifier, often necessitating an expert in the loop. She highlighted the challenge of systems without ground truth, such as sales outreach emails or self-driving near-misses, where safety engineers might have conflicting opinions on outcomes. Zhang also introduced the concept of 'building the harness' to manage the tension between unleashing an agent's probabilistic nature and imposing deterministic workflows.

G2 has implemented a program called 'AI interns,' hiring individuals fresh out of college or high school to perform AI-driven tasks. Zhang cited a specific example where an AI intern, an incoming Stanford freshman, completed a task in eight hours that previously took the engineering team three days. This task involved running runoff emails. Zhang observed that this new generation of technology workers is native to AI, which is helping to close the skill gap between those building cutting-edge systems and those still ramping up. She noted that while the gap between expert and novice users is currently visible, it is expected to narrow over time as AI becomes more integrated into basic training.