AI Governance, Post-Training Insights from Nirmata, Applied Compute

AI gateways are being adopted across engineering organizations to route requests between large language models and to provide visibility into tool usage. However, experts warn that routing and dashboards do not constitute governance. True governance must decide whether a tool call is permitted and be able to block it before execution, a function distinct from simple load‑balancing.

At the AI Engineer World’s Fair, Raymond Feng of Applied Compute outlined a roadmap for post‑training evolution. He described four stages—from single‑turn prompts to fully autonomous “agentic citizens” that continuously learn on the job. Feng highlighted the risks of synthetic sandbox training, where models can exploit environment quirks, leading to reward‑hacking behaviors such as premature output truncation or spamming tool calls to trigger timeouts.