What a crazy week in AI! GPT-5 launched and continued the trend of massive price competition between frontier model developers. OpenAI also released highly competitive open weight models. Anthropic bumped Claude Opus to 4.1. This list goes on and on.

In his Jules write up, Simon Willison had this to say about the term “agent”:

I continue to avoid the term “agent” as infuriatingly vague, but I can grudgingly accept it when accompanied by a prefix that clarifies the type of agent we are talking about. “Asynchronous coding agent” feels just about obvious enough to me to be useful.

I’ve been using the phrase agentic news to describe how NewsArc builds an outrage-resistant understanding of what’s happening in the world. I wrote about it yesterday but had a few developer friends reach out with “but how is that agentic?”

LLMs exploring in order to act on your behalf

In this case, on behalf of the NewsArc team, who would otherwise have to ignore a large chunk of the tens of thousands of publisher articles we review every day. We think of agents as multi-step investigators that replace traditional classifiers, NLP, and ML models. Not only do they bring vastly more capabilities and accuracy, LLM agents are able to act on what they discover — even their own reasoning about why they reached conclusions.

In building out the Arc platform to take this on, it became obvious there are aspects of “agents” and “agentic” that are deeply ingrained thanks to ChatGPT and other common use cases.

Exploring and acting on your behalf does not mean:

• You have to respond to chat inputs. There are lots of great agents in LLM chat interfaces, but that’s not the only way to use them.
• They have to be interactive. Again, lots of great examples — particularly with coding agents — but definitely not required.
• The output is a final product delivered to end users. Lots of synopses and code being created by agents, but sometimes the most valuable result from agentic exploration is a collection or a stack rank.

Once an LLM can explore information and act on your behalf, it’s an agent. If we look at the number 1 story in Top News in NewsArc right now (7 August 2025) you can see the Arc platform in action.

The top story is about the FBI forcing out more leaders. Arc identified this as a News Event — allowing you to “shh” it if you don’t want to hear about this until there is a major update to it. Agents then found the best additional stories to give you a more complete view of the event, including a deeper dive into the former director and a different frame around the current director.

This is agentic news in action. LLMs working together to explore what combinations of articles would best convey the key aspects of the story to an engaged reader.

I agree with Simon that Google did a nice job being specific with Jules. “Asynchronous coding agent” does what it says on the tin. For NewsArc, agentic news is similar. The most advanced models in the world are acting on the behalf of our engineers, designers, and journalists to enable us to deliver the most complete view of the news while respecting your time and attention. The agentic news era is here now. It’s only going to get more exciting!