Technical writing on Agentic AI, Reinforcement Learning, LLM alignment, and production ML engineering. Published here — not on Medium, not on Substack.
While exploring multi-agent systems, I realized the coordination patterns we use in AI today aren't entirely new — similar patterns have existed for decades in organizations, governments, biology, and economics. This article maps multi-agent architectures onto the Complexity Theory of Organizations.
Standard RL agents collapse the moment market regimes shift. Meta-RL — training agents to learn quickly, not just to act well — offers a principled solution. Here's the architecture and the evidence.
Reinforcement Learning from Human Feedback got us to GPT-4. But its fundamental bottleneck is human annotation speed. Here's why Reinforcement Learning from Agent Feedback is the logical successor — and what it means architecturally.
A deep technical breakdown of what actually separates a real AI agent from a wrapper around an LLM — covering memory, tool use, planning, and self-reflection.