The Good, The Bad, And The Future Of Ai Agents

Hey there, and invited to Decoder! I’m Hayden Field, elder AI newsman astatine The Verge and your Thursday section impermanent host. I’ll beryllium subbing successful for Nilay for a mates much episodes, and I’m excited to support diving into nan good, nan bad, and nan questionable successful nan AI industry. 

Today, I’m talking pinch David Hershey, who leads nan applied AI squad astatine Anthropic. David useful pinch startups to thief them fig retired really to champion use Anthropic’s tech, positive testing caller AI models to understand their limits.

I wanted to person David connected because Anthropic released a new AI model called Claude Sonnet 4.5 earlier this week, and it’s been making waves. (For reference, Claude is to Anthropic what ChatGPT is to OpenAI.)

The caller model, Sonnet 4.5, is being billed arsenic a large breakthrough successful autonomous, agentic AI, particularly for coding purposes. These types of AI products can, successful theory, beryllium fixed analyzable tasks and past spell disconnected and complete them complete nan people of galore hours aliases moreover aggregate days. Anthropic says this peculiar exemplary tin tally for up to 30 hours consecutive without immoderate quality involution — each while moving connected a singular task, for illustration building a package exertion from scratch. 

For nan past twelvemonth aliases so, companies for illustration Anthropic, Microsoft, OpenAI, and much person been promising that this agentic exertion would beryllium nan adjacent shape of AI, nan adjacent large hype-filled point that comes aft general-purpose chatbots. They opportunity it could really unlock generative AI’s potential, and it’s existent that they’ve made immoderate strides. 

But arsenic we’ve seen truthful far, agents aren’t rather location yet, and they person a ways to go. Most of america are not, successful fact, sending agents disconnected connected nan net to do our bidding, and we’re surely not giving them tasks that mightiness return 12, 24, aliases moreover 30–plus hours of autonomous activity without quality handholding. At least, not yet. 

At nan aforesaid time, galore companies are looking astatine agents arsenic nan breakthrough that’s expected to unlock immense productivity gains from AI models, including nan opportunity to usage them to switch aliases augment quality labor. 

So I wanted to beryllium down pinch David, who spends a batch of clip testing retired what modes for illustration Claude Sonnet 4.5 tin and can’t do, to inquire him wherever we are connected this committedness of AI agents. I wanted to talk astir what these types of products are bully astatine from a user standpoint, beyond programming purposes, and besides what nan way guardant looks for illustration arsenic agentic exertion progresses.

If you’d for illustration to publication much connected what we talked astir successful this episode, cheque retired nan links below:

• Anthropic releases Claude Sonnet 4.5 successful latest bid for AI agents | The Verge
• ChatGPT’s built-in Buy Now fastener has arrived | The Verge
• OpenAI really, really wants you to commencement your time pinch ChatGPT Pulse | The Verge
• Anthropic’s Claude AI is playing Pokémon | The Verge 
• AI agents are subject fabrication not yet fresh for primetime | The Verge
• Agents are nan early AI companies committedness — and desperately request | The Verge 
• Amazon is betting connected agents to triumph nan AI title | Decoder

Questions aliases comments astir this episode? Hit america up astatine decoder@theverge.com. We really do publication each email!