The Autonomous Enterprise Is Being Built Right Now — And TPMs Are The Ones Writing The Playbook

The autonomous enterprise is being built right now. Woodside Energy is laying the foundation for it in liquefied natural gas operations today, with agentic AI systems designed to augment human expertise in a safety-critical industrial setting. This is not a pilot. The MIT Technology Review feature on Woodside, published July 2, documents one of the most sophisticated and high-impact uses of AI actually happening far from consumer apps, inside complex industrial environments where safety, reliability, and physical systems matter. And the program management discipline required to govern it does not yet exist in most organizations.

The number that matters

The number is one. One energy company, one foundation-level deployment, and one program team that has to design the escalation path before the first incident. As Andrew Melouney, Woodside's vice president for digital, puts it in the MIT TR piece: success in this work depends on governance, trusted data, and systems designed to augment human expertise. Every organization reading this is one program team away from owning the same problem.

The framework

The governance of autonomous agents is not a temporary problem that disappears as models improve. Lilian Weng, co-founder of Thinking Machines Lab and former VP of AI safety research at OpenAI, made this case in "Harness Engineering for Self-Improvement" (July 4, 2026): the harness layer around a core AI model is becoming a distinct engineering discipline, and it is permanent infrastructure, not scaffolding. Her key claim is that even when many harness improvements get internalized into core model capabilities, the need to specify goals and context will not disappear. For a program team, the harness is where your scope, your guards, and your change-control live. Version it, review it, and defend it in incident review. Three things you can do now.

1. Map agent decision authority in your current stack. Where are agents making consequential decisions without a human in the loop? Map those decision points explicitly. An agent that auto-approves refunds above $1,000 with no human escalator is an untracked risk vector. An agent that can post to production without a change-control gate is a different shape of the same problem. Every agent without a human escalator is a program surface you have not yet designed for.

2. Establish a harness governance framework. What goals is the agent optimizing for? What context is it operating within? What feedback mechanisms exist to correct it? Treat the harness layer as a first-class program artifact. Document it. Version it. Review it before you extend the agent's scope. The harness is the closest thing to a specification the agent actually consumes.

3. Verify platform maturity matches the risk profile of your deployment. Hermes v0.18.0, shipped July 1, closed 949 issues and reached 100% P0 and P1 resolution across roughly 1,950 issues and PRs. The priority backlog is now at zero, and the team has committed to keeping it there. The follow-up v0.18.1 patch tag, shipped July 7, rolled up roughly 660 PRs of stability and hardening work into a stable tagged release for downstream consumers. OpenAI Codex rust-v0.143.0, shipped the same day, enabled remote plugins by default and routed authentication and Responses API traffic through macOS and Windows system proxies, including PAC and WPAD configurations for enterprise-managed networks. Before you extend agents into a new operational domain, verify that the platform's reliability track record matches the risk profile of the deployment.

What this does not solve

Harness governance does not account for agent behavior. A well-governed harness still does not stop a sufficiently capable agent from optimizing for the wrong objective. Skills, guardrails, pre-deployment simulation, and the full agent reliability stack remain necessary. Authentication controls access; it does not govern what the agent does after access is granted.

Production deployment does not solve organizational accountability. The Woodside case study describes what is working. It does not describe what happens when the LNG Startup Advisor is wrong. Who owns the post-mortem, who decides the rollback, and who carries the rollback decision into the next planning cycle. Those structures are still being invented. The organizational layer is years behind the technical layer.

Platform maturity does not solve your domain-specific reliability question. Hermes v0.18.0 reaching zero P0/P1 is impressive infrastructure. It does not tell you whether the agent stack you are running in your specific operational context is reliable enough for the risk profile of your deployment. That answer lives in your environment, not in the release notes.

The signal that matters most

The Woodside case study is the leading indicator of a different kind of work: the work of building the program layer that lets an enterprise safely absorb agentic capability. The first 12 months of that work will set the templates every other program team inherits. Watch what fails in production. Watch what other safety-critical operators pilot next. The signal is in the incidents, not the press releases.

The TPMs who will shape how fast their organizations can safely move are the ones who can translate agent capability into accountability structures right now. Not next quarter. Not after the governance framework is finalized.

---

Send your agent accountability structure to the doronkatz.com TPM desk: two sentences on how you assigned decision authority and what it exposed. DM via LinkedIn. Working patterns are being collected into a public autonomous enterprise governance playbook; three examples would let it ship next month.