Thoughts on agent workflow investigation, AI observability, and building the neutral layer that explains what matters.
Vendors cap liability at 12 months of fees. Insurers are adding AI exclusions. Regulators are writing new rules. Most enterprise contracts still treat AI agents like passive software.
Every vertical AI product sits on top of the same nine-layer infrastructure stack. Most teams only see their slice. This is the complete map — from silicon to investigation — with the interactive diagram covering ~200 vendors.
Jensen Huang says every SaaS company will become a GaaS company — agentic as a service. The frameworks to build agents exist. The tools to investigate what they're actually doing don't.
Replit deleted a database. Cursor wiped PocketOS in 9 seconds. Air Canada's chatbot made a binding legal promise. Every dashboard was green.
1,348 documented cases of AI-hallucinated legal citations worldwide. The problem didn't stay with solo practitioners — it reached the most prestigious firms.
Traces show events. APMs show spans. Neither explains whether the workflow was correct, allowed, efficient, safe, or worth changing.
When agents cite evidence they never retrieved, traditional observability says "run completed." We think that's not enough.
Harvey cares about correctness. Decagon cares about refund risk. Cursor cares about unsafe edits. One dashboard can't serve all three.