AMESA launched AutonomyStack, a platform to build industrial AI agents, train them in simulated operations, and test them against operating KPIs before they touch a live process. Human experts set how much autonomy each agent gets and when it must escalate, and an integration layer connects agents to plant systems through APIs and MCP. The release names no customer or result yet.
DOE made a conditional commitment of up to $5 billion to finance uprates and modernization at eight Constellation nuclear plants in Illinois, Maryland and Pennsylvania, adding nearly 1.1 GW. Constellation still has to meet technical, legal, environmental and financial conditions before any money moves. Getting more out of reactors already running is the fastest firm power available.
OpenShell runs each AI agent in an isolated sandbox where a policy decides every file, system call and network connection it can make. Agents never see real credentials; OpenShell adds them only to requests bound for approved endpoints, and risky policy changes wait for human review. That is the kind of hard fence agents need before they go near plant networks.
For IT and OT security leads: piloting agents inside plant or utility networks.
Found via @RoundtableSpace on X
An MIT-licensed Python reference turns telemetry, alarms, an asset ontology and maintenance knowledge into answers tied to a specific asset and alarm, with guardrails that block ungrounded numbers and unsafe writes. It scores those answers with a repeatable eval harness. It says plainly that operators want which asset and which alarm, not a fluent paragraph.
For reliability and controls engineers: testing an AI assistant on plant data.