In April, Rhombus quietly shipped something I didn't expect to see this soon: a plugin that teaches Claude Code, Anthropic's terminal agent, how to operate their camera platform. One CLI command snapshots your entire deployment, every site, every camera, a still image from each, into files the agent reads before it answers anything. In July they published the workflow that made me sit up: the agent checks occupancy from cameras after 5 PM, and for any floor empty more than an hour, it calls the building management API and drops the HVAC into energy-saving mode.

I spent this week reading their GitHub repos and the actual skill files, not the press release, and here's what I understand so far. They didn't bolt a chatbot onto a dashboard. They gave the agent what you'd give a new employee: a map of the site, documentation for the tools, credentials, and permission to work. For twenty years the user of a security platform was a human at a console. This is the first vendor I've seen design for a different user, an agent with a terminal. "List every offline camera across my deployment and group them by location" is now a complete query. No console, no filters, no CSV export. That's genuinely impressive work, and the pattern is public, so others will follow.

And the more I studied it, the more questions I ended up with. I don't have answers to these. I'm sharing them because I suspect the rest of the industry will hit them too, and I'd rather we start talking about them now.

The first one comes from my old life. I spent 21 years in law enforcement, and I sat through enough evidence hearings to know that the record of how evidence was handled matters as much as the evidence. So: when an agent pulls frames from an alert clip, describes what it saw, and stitches a multi-camera incident video, what is the record of what it did? Which footage it touched, what it was asked, what it decided, what it skipped. As far as I can tell, nothing like that exists yet, anywhere in the industry. If defense counsel asks who reviewed the footage and the answer is "a language model with an API key," what happens next? I honestly don't know. That question keeps me up.

The second one is about identity. In July, Genetec disclosed a flaw in Security Center where live video could be retrieved without proper authentication checks. Credit to them for the fast disclosure and hotfix. What struck me was something more basic: everything I know about identity in our platforms assumes the client is a person who logs in. What does scoping look like when hundreds of agents hold API keys and work around the clock? I'm still learning how the IT world handles this, and I'm not sure their answers transfer cleanly to systems that watch physical space.

Third, every real deployment I've ever walked is a mixed fleet. The agent tooling shipping today stops at one vendor's boundary. How does this pattern extend past that line? Open question.

And the environment around these questions is moving. EU AI Act transparency duties for emotion-recognition and biometric-categorization systems went live August 2. LAPD let its Flock contract expire this summer over data ownership concerns, and a Colorado county switched ALPR vendors citing trust. I don't read those as anti-AI signals. I read them as the public asking the same kind of questions I'm asking here: not "does it work" but "how do we know what it did."

I don't have a framework to sell you and I'm not warning anyone off. The occupancy demo is real, the efficiency is real, and I think this is the most interesting thing to happen in our industry since cameras moved to IP. I just think the industry that ends up explaining itself in court has to learn this technology more carefully than most, and I'd rather learn it together than watch us each hit these walls alone.

If you've thought about any of this, agree or disagree, I want to hear it. Teach me something in the comments.