Agents stop being experiments and start being plumbing. Coinbase’s Agentic Wallets is the clearest example: an interface that lets agents hold value and act on it turns models into actors in financial systems. The technical move is obvious; the real move is organizational — teams must treat agents as infrastructure, not features. Pay attention to Principle 07 — Build the Island and Principle 09 — Agentic Coordination is a New Org: environments, sandboxes, and runtime topologies (see hive and OpenAI’s harness work) are where trust and velocity are won.

The governance and security question follows immediately. OpenAI’s path to giving the US military access via GenAI.mil and the Pentagon’s pressure to loosen restrictions on classified networks (Reuters) show the same tension: autonomy at scale versus controllability under law. Coinbase’s agentic wallets amplify that tension — money plus autonomy is a governance problem, not just an engineering one. Protect the product by hardening Principle 10 — The Law, closing the Gate (Principle 15 — The Gate), and building an internal immune system (Principle 14 — The Immune System) like the one described in The Information’s report on OpenAI’s monitoring tool.

Tooling and open weights lower the bar for outcome-oriented teams. Z.ai’s GLM-5 (and Simon Willison’s write-up, GLM-5: From Vibe Coding to Agentic Engineering) plus OpenAI’s Codex-driven playbook (Harness engineering) and inline Skills make it plausible for small teams to ship agentic workflows quickly. That is a win for Principle 08 — Ship the Artifacts and Principle 06 — Legible Landscapes: when models, artifacts, and runtimes are composable and observable, outcomes become auditable. The New York Times’ custom LLM pipeline for podcast signals (Manosphere Report) is a concrete, practitioner-level reminder — legibility is how journalists, engineers, and product owners actually trust agent outputs.

What practitioners must do now is straightforward and urgent. First, design execution environments and skills as first-class artifacts — treat them like product components and version them. Second, instrument the Gate and the Immune System: policy, access controls, and internal detection must be as automated and observable as the agents themselves. Third, favor open, composable models and runtime patterns so your team can iterate on outcomes instead of wrestling with brittle integrations — learn from GLM-5, Codex harnessing, and the hive project. Outcome engineering demands this roster of moves: build the island, ship the artifacts, and harden the gate.