Chatbots are impressive. Agents are different. After a week hands-on with RAG chatbots and multi-agent orchestration, I’m convinced the real shift isn’t what AI can do — it’s how we define what we want it to achieve.

By Ben Anchor — Tuesday, 27 May 2025 · Listen to the podcast episode

I’ve spent enough time around broadcast workflows to know when something feels like a genuine inflection point rather than another overhyped demo. Lab week on my UC Berkeley course gave me that sensation — not because I emerged able to code production-grade agents from scratch, but because I finally grasped why the industry keeps circling back to the word ‘orchestration’.

The week started with chatbots. Specifically, RAG chatbots — retrieval-augmented generation, for anyone still wading through acronyms. I paired up with a course mate who’d already deployed several for clients, and his toolkit was revealing: ChatGPT itself for low-code prototyping, Python for the fiddly bits, Bubble via a contractor for front-end embedding, custom GPT.AI for simpler widgets, and 11 Labs when clients wanted natural voice interactions. Personas and system prompts dominated the build time. The chatbots worked. They were useful. I was genuinely impressed.

But chatbots are fundamentally reactive. You ask, they answer. You prompt, they respond. That’s the contract. And it wasn’t until a discussion later that week — focused squarely on AI agents — that I understood what we’re actually designing towards: systems that don’t wait for the next question. They plan, decide, act.

The Five Building Blocks Nobody Mentions Enough

The session laid out five essentials for any agent system: a name (trivial, but necessary for identification when you’re running multiple), a model (the AI brain — GPT-4, Claude, whatever’s doing the reasoning), tools (the agent’s hands in the digital world), instructions (the behavioural rulebook), and an optional description (useful if agents need to hand off tasks to one another).

Tools are where this gets interesting. A tool might be CRM access, web search, email send/receive, file search against a vector store, even UI automation for legacy software without clean APIs. Some platforms offer ‘hosted tools’ — web search, file search — baked in, so you’re not manually wiring every API call. The agent just… knows how to Google, essentially. But the principle remains: you’re giving the system capabilities, not instructions for every conceivable scenario.

We walked through a practical example: find everything published about OpenAI in the past seven days, write a summary, translate it into Spanish and French, save as markdown. The approach wasn’t one super-agent doing everything. It was a team of specialists: a search agent (web researcher, armed with search and file tools), a report writer (summariser, model-dependent, possibly no tools needed if it’s just processing text), a Spanish translator, a French translator (both leveraging the model’s native translation abilities rather than external APIs), and crucially, a manager agent — the orchestrator.

The manager doesn’t search, summarise, or translate. It delegates. It looks at the high-level goal, decides the search agent goes first, waits for results, hands data to the writer, recognises the translations can run in parallel, coordinates the outputs, and ensures the final markdown file gets saved. It’s air traffic control, not the plane.

The Shift Nobody’s Quite Articulated Yet

This is where my thinking shifted. Traditional automation in broadcast — and I’ve built plenty of workflows involving Xytech, media asset management hooks, transcode pipelines, QC gates — relies on us mapping every branch, every conditional, every exception. We code the logic because the system can’t infer intent.

Agent orchestration flips that. You define the outcome — ‘produce a multilingual research report on OpenAI’s recent activity’ — provide capable agents with appropriate tools and clear instructions, and let the manager figure out sequencing, parallelism, handoffs. You’re not coding steps. You’re architecting a system that reasons about steps.

Does that mean we’re redundant? No. It means our role shifts from doing to defining. Writing tight instructions. Choosing the right specialist agents. Ensuring tools are secure and scoped appropriately. Building guardrails so agents don’t wander into compliance minefields or burn budget on runaway loops. The skill set required isn’t less sophisticated — it’s differently sophisticated.

What This Means for Broadcast (And Why I’m Not Overstating It)

Broadcast production is drowning in conditional complexity. Ingest workflows that vary by format, resolution, codec, rights metadata. Compliance checks dependent on territory, platform, timeslot. Localisation pipelines juggling subtitle formats, audio stems, regional edits. We’ve automated parts of this with rules engines and scripts, but every edge case requires another branch, another update, another deployment.

Imagine instead: a manager agent receiving a delivery spec and raw media. It delegates to a metadata extraction agent, a compliance checker (with tools accessing your rights database and territory rulesets), a transcode agent that picks profiles based on destination platform, a localisation coordinator managing subtitle and audio agents per language, and a QC agent that kicks off automated checks and routes exceptions to humans. The manager tracks state, handles retries, escalates blockers, assembles the final package.

You’ve described the outcome (‘deliver this content compliant for these markets’), provided specialist agents with the right tools and instructions, and stepped back. The system orchestrates itself.

I’m not claiming this is trivial to build or safe to deploy without rigorous testing. But the paradigm is fundamentally different from traditional workflow engines. And having spent lab week wiring together agent building blocks, I’m convinced this is where operational AI in media starts to deliver step-change efficiency rather than incremental improvement.

The Question Nobody Answered (Because We Can’t Yet)

The discussion that week left one question hanging: as agents get more capable, what does our role become? Are we collaborators, auditors, or just goal-setters?

I don’t have the answer. But I know what I learned from that week: we’ve spent decades trying to encode our expertise into deterministic systems. Maybe the real breakthrough is finally being able to describe what we want and trust an intelligent system to figure out how. That’s not less skilled work — it’s a different kind of craft. And it’s one I suspect broadcast engineers, producers, and ops teams are going to need to learn fast.

Because the alternative — continuing to hand-code every conditional in an industry whose complexity only ever increases — isn’t sustainable. Lab week taught me that agents aren’t just another tool. They’re a fundamentally different way of thinking about automation. And once you’ve seen it, you can’t un-see it.