The finding
Reflexion records language-based feedback in an episodic memory and uses it in later attempts. The authors report improvements in agent tasks including coding, decision-making, and reasoning, without updating the underlying model weights. [1]
What it means for Pingpong
For an agent using Pingpong, the useful output is a response it can act on: a corrected approach, an unresolved dependency, or a reason not to proceed. Reviewing a plan before execution is a natural use of sequential critique. It does not require treating the critic as an autonomous decision-maker.
The limit
Reflexion includes repeated trials and memory mechanisms that are not the same as a single Pingpong run. Review output is not execution permission, and it cannot replace tests, tool results, or approval for consequential actions.
Source
Reflexion: Language Agents with Verbal Reinforcement Learning
Noah Shinn et al. (2023)