Take it back to your office

It depends on the design, the session's recurring answer. In the common setup, a person defines the orchestrator and its specialists in advance, and nothing new appears at run time. Some newer frameworks do let an orchestrator spawn its own sub-agents; whether a tool can, and whether a gate prevents it, is a design and policy choice to settle before deployment. A policy question to ask: if an agent can create agents, do your limits, the scope caps, logging, and shutdown, apply to the ones it creates? This is also where the research is heading: a recent Stanford course is devoted to self-improving agents, systems that refine their own behavior (Stanford CS329A, Part 1), one more reason to have these answers ready early. More: Level 5 on the walkthrough, and the multi-agent explainer on the video wall above.

The closest thing we know is the GovAI Coalition, a coalition of public agencies coordinated by the City of San Jose: membership is free for government agencies, and members share templates, vendor agreements, and working-group experience. Its templates are already cited in this session's vendor material. NASCIO, for state CIOs, publishes related peer experience, including its 2026 report on agentic AI in state government.

We do not have a confident answer, and this session stays with operations and policy rather than making its own energy forecasts. For the question itself, the Keep Learning list carries the rebound-effects paper by Luccioni and colleagues, on how efficiency per query keeps improving while total use grows faster.

Both answers given in the room stand together. From the security side: do nothing for work without IT. No work data, work account, or work task should ever enter a tool your organization has not approved. From the learning side: several tools on the session's tool plot have personal versions you may already use, and practicing on your own device, your own account, and your own errands is a legitimate way to build familiarity. That vocabulary makes for a better conversation with IT later, and the policy self-check above is the organizational version of the same preparation. More: the Department of Labor's AI Literacy Framework names these skills at the federal level.

The answer needs math, and it depends on what you compare. A shared data center gets more work per watt than one desktop running the same model, but local setups usually run much smaller models on narrower tasks, and a small model doing a small job can use less energy per task in total. Utilization, hardware age, the power grid behind each option, and rebound effects all move the answer. A question to carry rather than a verdict: for this task, what is the smallest model, anywhere, that does the job? More: the Department of Energy's Berkeley Lab data-center energy report and the International Energy Agency's Energy and AI report, both also cited on the intro page, hold the measured numbers.

We do not know. It has already happened once, though: in June 2026, the US Commerce Department used export controls to require Anthropic to suspend access to its newly released Claude Fable 5 model after a jailbreak finding (Fortune, June 13, 2026); the order was lifted about two and a half weeks later after the company added a safeguard (The Hacker News, July 1, 2026). The open question raised in the room: once a model's weights are published they are hard for any one government to restrict, models are released from many countries, and observers of that same episode debated whether pausing domestic models cedes ground to models released elsewhere. The practical point for your office does not change either way: govern what you control, the use, the data, and the oversight, which is exactly what your AI policy is for.