Nilou’s Notes
What we’re learning about building agents that enterprises can trust, from Across AI’s CEO, Nilou Salehi, and the team.
Follow Nilou’s Notes on LinkedIn Latest · Safe by Design Guardrails catch most bad behavior, and “most” is the problem. Three architecture decisions that secure our agentic AI platform before any guardrail is needed. Utsav Kanani
- Regression blindness: self-improving agents have a problem Agents are much better at predicting what a change will fix than what it will break.
- Did your agent learn, or did it get lucky? An agent approved the same request three times and refused it the fourth. How do you know an agent learned a process rather than got lucky?
- Securing the Agentic Enterprise Security used to ask “Who are you?” Agentic systems force us to ask “Why are you doing this?”
- Why OpenClaw-Style Memory Breaks in the Enterprise When memory escapes its permissions. Agents become valuable when they can remember and connect dots across data sources, and dangerous at exactly the same time. Nilou Salehi and Afshin Nikzad
- Why “Just Use MCP” Breaks in the Enterprise Content is cheap now. In the enterprise, content without context is useless. What determines whether enterprise AI works is context integrity. Nilou Salehi and Jason Taylor
- Reasoning Graphs: Turn Context to Action Agents use Reasoning Graphs to do multi-step reasoning across enterprise data, store organizational memory, and operationalize with next-best-actions.