Today's stories sit at the intersection of AI governance and developer tooling. Anthropic is making visible structural commitments to outside accountability—adding a physicist with real scientific credentials to its oversight board and opening its internal systems to independent safety evaluators. On GitHub, the theme is infrastructure: tools for running fleets of coding agents and keeping context alive across sessions and tools. The underlying current in both threads is trust, just aimed at very different problems.
Physicist Christopher Stubbs Joins Anthropic's Long-Term Benefit Trust
Christopher Stubbs, an experimental physicist and former Harvard Dean of Science who helped discover the accelerating expansion of the universe, has joined Anthropic's Long-Term Benefit Trust. The Trust is an independent body with the authority to appoint a majority of Anthropic's board and advises on risks and societal impacts—giving it genuine structural leverage over the company's direction. Adding a scientist with Stubbs's background in experimental rigor signals that Anthropic wants the Trust to function as a substantive check, not a ceremonial one.
Anthropic Opens Its Systems to Independent Safety Evaluators
Anthropic has granted a six-person team from METR—a nonprofit focused on AI safety evaluation—access to its internal systems, hardware, and staff at roughly the same level as Anthropic's own risk teams, and METR will publish its findings publicly. The company is also working with SecureBio and Accenture under a three-tier access model, with more organizations in early talks. Independent external audits of frontier AI systems have been called for for years; this is one of the more concrete implementations of that idea from a major lab.
Orca Orchestrates a Fleet of Coding Agents in One Place
Orca is an open-source orchestrator that runs multiple AI coding agents—Claude Code, Codex, OpenCode, and others—simultaneously, each in its own git worktree, all tracked in one interface. It can fan a single prompt out to several agents, let you compare the results, and merge the best output into a single branch. Mobile companion support, SSH remote worktrees, and built-in GitHub and Linear integrations round it out into something closer to a full multi-agent workbench than a thin wrapper.
ai-memory Carries Agent Context Across Sessions and Tools
ai-memory is a Rust-based memory layer that persists context across sessions, AI tools, and machines. Lifecycle hooks silently record prompts, tool calls, and session boundaries, then write them as markdown pages to a git-backed wiki—so switching from Claude Code to Codex no longer means starting cold. The files stay human-readable and editable, no API key is required for the default path, and the same server scales from a single developer to a whole team.