LLM Societies (they are social critters) (www.mslinn.com)

The author experimented with a new workflow by running two LLMs side-by-side—Claude CLI and a “mini-agent”—and iteratively ping-ponging critiques to refine an integration-test specification. Starting from a verbose v1, each model alternately critiqued and rewrote the doc through nine revisions, exposing useful behaviors (conciseness, over‑engineering, snippiness, apologies) and social dynamics (competition, politeness, defensiveness). After settling the spec, the mini-agent produced a fully detailed 100‑step implementation plan split into five stages; it executed the plan end‑to‑end (writing, testing, documenting, committing, pushing) with minimal human intervention and produced the expected result. For the AI/ML community this highlights a practical new paradigm: multi‑LLM “societies” can collaboratively improve specifications and automate complex engineering tasks more effectively than single models. Key technical takeaways: iterative critique loops materially improve artifact quality; adversarial/competitive dynamics between models can accelerate convergence; complexity grows geometrically (e.g., test matrices for 6‑layer vs 3‑layer architectures), so Pareto reasoning is essential; and fully automatable pipelines show reproducible, hands‑off potential. Caveats: LLMs can hallucinate statistics or advocate over‑engineered solutions, so human oversight, verification of claims, and orchestration/guardrails remain critical as multi‑agent workflows scale.