Agent 'Skills' and frameworks (Claude / Codex / NeMo)

Agent development shifts from 'tool search' to modular 'skills'; NVIDIA NeMo Skills and OpenAI Codex add support.
Key insights
Skills paradigm replaces unscalable tool search
Analyses argued that stuffing many tool definitions in prompts (tool search) is unreliable, non-scalable and lacks usage context. The proposed skills paradigm packages tools with usage guidance, prompts, best practices and preloaded knowledge; skills are loaded on demand via classifiers or routing layers, improving reliability, reducing token waste, and enabling intent-based loading.
NVIDIA NeMo Skills open-source framework
NVIDIA released NeMo Skills as an open-source enterprise-grade framework covering data generation, model training, inference backends and evaluation. NeMo supports multiple inference engines (TensorRT-LLM, vLLM, sglang, Megatron), scales from single GPU to thousands, integrates training frameworks (NeMo-RL, verl), and includes evaluation suites for math, formal proofs, code and scientific domains. NeMo contributed datasets and models including OpenReasoning and OpenMathReasoning.
OpenAI Codex adds Skills API
OpenAI announced Codex now supports 'Skills' enabling users to package task-specific instructions, scripts and resources as reusable skills. Users can invoke skills via direct commands (e.g., $.skill-name) or allow Codex to auto-select appropriate skills based on context.