Blog

Notes on building OmniAtlas, AI engineering, and how we teach it.

All llm-agents (1) reflexion (1) reinforcement-learning (1) td-learning (1)

Reinforcement learning · LLM agents · Reflexion

Two Loops, One Value Function

TD learning Reflexion · Verbal RL · Value function

reinforcement-learning reflexion llm-agents td-learning

Two Loops, One Value Function: TD Learning Meets Reflexion

An LLM rewrites a value function's code from the agent's failures while TD learning fits the weights. In five iterations the mean reward went from -760 to about -116.

Hamid Ahmadian

Oct 19, 2026 · 18 min read