According to @karpathy on X, an 8-agent setup using 4 Claude and 4 Codex instances, each on a single GPU, failed to produce reliable gains while attempting to remove a logit softcap in nanochat without regression; the multi-agent research org tried configurations like 8 independent researchers and a chief-scientist model with juniors, but agents generated weak ideas and poor experiment hygiene (no strong baselines, ablations, or compute controls) despite being strong implementers of well-scoped tasks (as reported by Karpathy’s thread and video post on Feb 27, 2026). According to @karpathy, the orchestration used git branches per research program, feature branches per agent, git worktrees for isolation, simple file-based comms, tmux grid sessions, and no Docker or VMs, highlighting a lightweight but auditable workflow for AI automation. According to @karpathy, business takeaway: multi-agent LLM research orgs currently need human PI oversight for hypothesis generation and experimental rigor; near-term opportunities include building agentic RAG playbooks for baseline enforcement, automated ablation and FLOPs control, reproducibility checklists, and evaluation harnesses tailored to model training tweaks like logit caps. According to @karpathy, the approach reframes prompts, tools, and processes as “org code,” suggesting vendor opportunities in agent orchestration platforms, experiment-tracking integrations, and guardrailed research pipelines for enterprise ML teams.