IronClaw: Rust-based assistant that runs tools in isolated WASM sandboxes ships a Rust/WASM runtime that executes untrusted tools in isolated sandboxes while keeping data local and encrypted. That gives outcome engineers a concrete blueprint for secure tool execution and local-first agents — build the island and harden the immune system (Principles 07,14).

cloudrouter: Skill letting Claude Code/Codex spin up VMs and GPUs exposes an agent skill that can programmatically spin up cloud sandboxes, GPUs, run commands, and automate browsers from the CLI. Outcome engineers can use this pattern to automate ephemeral compute for reproducible agent runs and integrate infrastructure control directly into agent workflows (Principles 03,07).

Moltis — AI assistant with memory, tools, and self-extending skills releases a self-hosted assistant with long-term memory, sandboxed tools, local LLMs, and runtime self-extension. It models a composable, auditable agent architecture for teams that need offline control, reproducible artifacts, and verifiable memory handling (Principles 06,07,15).

GPT‑5.2 derives a new result in theoretical physics reports GPT-5.2 conjecturing and helping prove a new nonzero single-minus gluon tree amplitude, with the result confirmed analytically by authors. The case underscores agent-assisted discovery workflows and the imperative for verification pipelines, rigorous model checks, and outcome audits when agents generate domain-changing outputs (Principles 03,16,02).

Introducing Lockdown Mode and Elevated Risk labels in ChatGPT launches Lockdown Mode to restrict external interactions and adds “Elevated Risk” labels to flag high-risk capabilities and reduce prompt-injection attacks. Adopt similar access controls, capability labeling, and gate mechanisms to protect deployed agents and meet regulatory and organizational safety requirements (Principles 10,15).