Every broadcast vendor pitch I've sat through in the last six months has used the word "agent." Most of them are selling chatbots. The distinction sounds academic until you look at what it means for your architecture, your integration footprint, and your risk profile.

A chatbot is a request-response system. A user asks something. The system retrieves or generates an answer. The conversation closes. Whatever happens next depends entirely on the human who asked. If you unplugged every operator and left the chatbot running, it would sit there doing nothing.

An agent is a different shape of system entirely. It has a goal. It observes its environment. It takes actions to pursue that goal. It evaluates the results of those actions and decides what to do next. That loop runs continuously, whether or not a human is talking to it.

In a broadcast context the difference is stark.

A chatbot answers "what's the status of encoder 4?" You ask, it looks, it tells you. Useful. Essentially a better search interface over your existing telemetry.

An agent continuously monitors encoder 4 alongside the rest of your signal chain. It detects that the output bitrate is degrading. It correlates that signal with upstream network latency, with scheduled maintenance activity, and with the fact that a live programme is going to air in 40 minutes. It concludes this combination of conditions is going to cause a problem. It escalates to the right on-call engineer with the context they need to act. None of that required a human prompt.

The gap between those two systems is not a feature difference. It's an architectural category difference. And it changes everything downstream of the buy decision.

A chatbot can usually be deployed as a thin layer over existing systems. It reads. It responds. Integration is one-directional. The risk surface is small because it can't do anything.

An agent needs to observe continuously, which means deep integration with monitoring stack and telemetry. It needs to act, which means write access to ticketing, notification, escalation, and potentially control systems. It needs state, because the decisions it makes depend on what it has already seen and done. It needs observability of its own reasoning, because when something goes wrong you need to audit why.

That's a fundamentally bigger build. It's also a fundamentally more valuable system.

The test when a vendor says "agent": does it do anything when nobody is talking to it? If the answer is no, you're being sold a chatbot. That's still a useful product. It's just not what's going to transform your operation.

Buy what you actually need. Don't pay for the word.

"AI-powered." "AI-native." "Intelligent." "Smart." "Cognitive." "Autonomous."

These terms mean almost nothing. They persist because they're hard to challenge in a meeting where most people don't want to admit they don't fully understand what the vendor is claiming. They're the broadcast AI equivalent of calling a stereo system "audiophile-grade."

The words that actually tell you something:

If a vendor answers those four questions crisply, you're talking to people who have actually built something. If they deflect, mumble, or default back to marketing adjectives, you're being sold a demo that hasn't been battle-tested.

Here's a useful exercise for your next vendor meeting. Ban the marketing vocabulary for 10 minutes. No AI. No agent. No intelligent. No smart. Ask them to describe what their system does using only concrete verbs. Monitors. Sends. Reads. Writes. Decides. Escalates.

You'll learn more in those 10 minutes than in the previous hour of slideware.

Vendors who publish their agent's scope of autonomy in writing. It's a strong signal they've actually thought about the boundaries.

"Multi-agent" as a feature. In most implementations it's one agent with some subroutines dressed up as a team. Ask to see the communication protocol between agents. If there isn't one, there aren't multiple agents.

Anthropic's framework for developing safe and trustworthy agents -- particularly the section on the balance between autonomy and human oversight. A useful lens for reading vendor claims.
https://www.anthropic.com/research/trustworthy-agents

The Clean Feed is published every Thursday. Forward this to someone who builds broadcast systems.