An OpenAI Codex developer has warned that deploying more than two parallel sub-agents typically burns tokens without improving output quality, citing what he calls a “coordination tax” where agents distrust each other and duplicate work (The Decoder). The warning aligns with broader token economics concerns; OpenRouter data shows weekly token consumption surged over 25,000 percent since January 2025, driven partly by unoptimized agent deployments and reasoning models (The Decoder).
Google Research introduced Retrieve-for-Train (R4T), a framework that trains a fan-out language model with reinforcement learning using groundedness, diversity, and alignment rewards, then synthesizes training data for a 53.9M-parameter diffusion retriever achieving 12- to 20-fold faster query fan-out (MarkTechPost). Separately, Knowledgator released GLiFormer, a 575M-parameter encoder scoring 91.10 F1 on nested JSON extraction without token generation, approaching GPT-5.6-luna performance while grounding values in source spans (MarkTechPost).