In the never-ending smorgasbord of AI hype, “agents” represent practical and worthwhile potential. AI agents are autonomous AI programs that can understand some context and take action in that context. Agents can autonomously perform a task that involves mapping a goal to its context and parameters (even if they’re not explicitly laid out), process data across multiple formats and ontologies to understand the goal and work through the task, call multiple functions across multiple apps, and take some action to achieve the goal. Unfortunately, however, while many are talking about AI agents, few are promoting actual products at CES.

Basically they’re software programs who “get stuff done” in our messy world without explicitly being programmed step-by-step like regular software. AI agents will “figure things out,” up to a certain point (as far as the knowledge graph lights the way).

For example, a very simple AI agent could listen in on a call where you say you want to go out for Italian food, then find out which one is the best rated in your area, book the reservation, and make an entry for it in your calendar.

The startup Rabbit, which made a splash at CES last year, famously made a courageous attempt at a multi-agent product, the r1. Alas, they were too ambitious, had raised too much money too fast, and were working on too short of a timeline to be successful building agents in a million different verticals on a 10-month turnaround.

Agentic AI has been around for a long time, but it’s been through an aggressive renaissance in the past 12 months when people started hooking up LLMs to knowledge graphs to “do things.”

Agents are also potentially critical in the media and entertainment industry, especially in production (death by a thousand spreadsheet tabs) and post-production (orchestrate many small autonomous tasks into one hybrid human-machine pipeline).

So agents are big in AI right now, which means they’re big et CES. Right?

Nvidia Goes “All-In” on Agents

Surprisingly, we’ve looked at a lot of products in the past few days, and AI agents are almost nowhere to be seen. Except of course in the CES keynotes and panels.

Agents were a big part of Nvidia CEO Jensen Huang’s keynote on Monday, and this man absolutely knows what he’s talking about. Agents are, in his words, probably “a multi-trillion-dollar opportunity” (or at least will be at some point in the future), and Nvidia is building impressive ML ops platforms to help with their deployment.

“The IT department of every company is going to be the HR department of AI agents in the future,” said Huang, referring to the agents’ orchestration layer, the “AI minions’ command center,” which would look a lot like if Retrieval-Assisted Generation was a John Waters musical.

Here’s the rub: meaningful AI agents are incredibly hard to do. Chiefly because they rely on context, and AI doesn’t do context very well. The wider the context, the more edge cases, the more uncertainty or ambiguity the agent has to navigate, the more uncertain the outcome.

On the “impossible” end of the AI Mission: Possible spectrum, autonomous vehicles are AI agents. And look at how long and how much money it has taken to get them on the road.

Agents of Travel: Delta Concierge and Glib

The universal rule of AI so far is that the narrower the context, the easier it is to deploy intelligence in a meaningful and trustworthy manner. There are less edge cases and it’s easier to build compelling products by focusing on one vertical.

The sweet spot with agents is travel. And it makes sense: booking travel is both a coherent and bounded semantic domain … and an annoying multi-step and multi-app process. So it’s no surprise that the two agentic systems we saw so far at CES have to do with travel.

In its flashy keynote at Sphere, Delta Air Lines rolled out Delta Concierge, an agentic-ish chatbot (strangely not formally labeled as an agent) that will be rolled out in the near future as part of its Fly Delta app.

According to the impressive demo, Concierge is capable of contextually handling many aspects of the travel experience, down to the itinerary to the airport and the Uber (or Joby, if you’re around when we finally have actual flying cars in cities) reservation. The app looks slick, it makes sense, and it feels right from a technical scoping standpoint.

Another clever agent-for-travel from French startup Glib caught our eye over at The Venetian. We’re rooting for Glib because its AI agent-for-travel play is both ambitious and spot-on from a product standpoint.

The basis of Glib is simple: enter your budget, your destination, your timeline, and Glib will figure out your flights, your hotel, and some cool things to do where you’re going. We tried it and it actually works pretty well. The app will also keep you up to date about museum closures, delays, etc. Of course a lot of polishing is needed, but this is an impressive start.

It usually takes 12 months for a tech to jump from the panels to actual products on the floor, so we expect a lot more AI agents next year, but to realize Jensen Huang’s Godzilla-size market opportunity the tech world will need less Rabbit r1s and more of the kind of product excellence and scrappy AI cleverness as shown in these two examples above.