Custom Kernels for All from Codex and Claude. Hugging Face ships a ‘cuda-kernels’ agent skill that generates production-grade CUDA kernels, hooks them into PyTorch, benchmarks on H100, and publishes artifacts to the Hub. Outcome engineers can now treat agents as creators of deployable GPU code — plan for artifact provenance, CI for generated kernels, and runtime verification (Principle 08, 16).

IronClaw: Rust-based assistant that runs tools in isolated WASM sandboxes. IronClaw executes untrusted tools inside Rust-backed WASM sandboxes while keeping data local and encrypted. This gives a concrete, local-first pattern for isolating agent-executed tools and reducing trust surface — adopt sandboxed runtimes and strict I/O gates when composing agent skills (Principle 07, 14).

cloudrouter: Skill letting Claude Code/Codex spin up VMs and GPUs. cloudrouter adds an agent skill to provision cloud sandboxes and GPU instances, run commands, and automate browsers from the CLI. Treat provisioning as an agent primitive: you must design access controls, cost metering, and reproducible sandbox images when agents can create infrastructure (Principle 07).

Moltis — AI assistant with memory, tools, and self-extending skills. Moltis offers a self-hosted assistant combining long-term memory, sandboxed tools, local LLMs, and runtime self-extension. Use this as a reference architecture for persistent, extensible agents — design clear extension APIs, capability gating, and memory hygiene to avoid drift and unsafe self-modification (Principle 06, 07).

GPT‑5.2 derives a new result in theoretical physics. GPT-5.2 conjectures a new scattering-amplitude result and an internal model helped prove it, with authors confirming the analytic proof. This demonstrates agents proposing novel claims that require rigorous verification — build audit logs, reproducible artifacts, and human-in-the-loop validation pipelines for agent-driven research (Principle 16).