Custom Kernels for All from Codex and Claude demonstrates agents generating production-grade CUDA kernels, integrating with PyTorch, benchmarking on H100, and publishing kernels to the Hub. This gives outcome engineers a concrete pattern for agents to produce deployable artifacts and speeds the pathway from intent to shipped GPU code (Principles 06, 08).

cloudrouter: Skill letting Claude Code/Codex spin up VMs and GPUs introduces a CLI agent skill that spins up cloud sandboxes and GPU instances, runs commands, and automates browsers. That moves environment provisioning into agent workflows and forces teams to design orchestration, ephemeral resource controls, and gating for safe execution (Principles 07, 09).

IronClaw: Rust-based assistant that runs tools in isolated WASM sandboxes presents a local-first runtime that executes untrusted tools inside Rust-backed WASM sandboxes while keeping data encrypted. Outcome engineers can adopt this as a pattern for secure tool execution and containment when agents need to run third-party skills or user-contributed code (Principles 07, 14).

Moltis — AI assistant with memory, tools, and self-extending skills ships a self-hosted assistant combining long-term memory, sandboxed tools, local LLMs, and runtime self-extension. It highlights the need for versioned artifacts, gates, and auditability so teams can safely allow agents to extend their own capabilities without losing control (Principles 07, 15, 08).

GPT‑5.2 derives a new result in theoretical physics reports GPT-5.2 conjecturing and contributing to a provable new amplitude result confirmed by authors. That underscores agent-assisted discovery’s power and the imperative for reproducible verification pipelines, provenance tracking, and outcome audits before trusting model-generated conclusions (Principles 03, 16).