At this point, I think we can safely call the period between Dreamforce ‘24 and Dreamforce ‘25 the ‘Year of Agentforce’. Last year, it was officially announced, and this October, we saw its latest iteration with Agentforce 360 – with many other announcements from Salesforce centering around its agentic AI product.
It’s very much front-and-center of Salesforce’s messaging at the moment, and the message from the CRM giant is clear: We are now in the age of the ‘Agentic Enterprise’.
But is this actually the case? Co-founder of ChatGPT creator OpenAI, Andrej Karpathy, recently spoke with host of the Dwarkesh Podcast, Dwarkesh Patel, arguing that we are not – as some have suggested – in the “year of agents”, but rather the “decade of agents”, saying there was still a lot of work to be done. Let’s take a look at who has the stronger case.
The Biggest Challenge for Agent Creators? ‘Making Them Work’
Karpathy later went on to clarify his position in a lengthy X post. Under the ‘On LLM agents’ subheading, he writes that his critique of the industry relates to what he describes as “overshooting the tooling” with respect to present capability.
He adds: “I live in what I view as an intermediate world where I want to collaborate with LLMs and where our pros/cons are matched up. The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code, and humans are useless. For example, I don’t want an agent that goes off for 20 minutes and comes back with 1,000 lines of code. I certainly don’t feel ready to supervise a team of 10 of them.”
Karpathy goes on to say he would like to “go in chunks” that he can keep in his head, where an LLM explains the code that it is writing, and can prove that what it did is correct, pulling the API documents and showing that it used things correctly.
He writes: “I want it to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I’m told works.
“I just think the tools should be more realistic [with reference to] their capability and how they fit into the industry today, and I fear that if this isn’t done well, we might end up with mountains of slop accumulating across software, and an increase in vulnerabilities, security breaches, etc.”
His opinion might make difficult reading for Salesforce Founder and CEO Marc Benioff, who very much appears to believe that AI agents are here and ready to deploy.
When grilled at Dreamforce ‘25 about the apparently low adoption of Salesforce’s agentic AI product, Agentforce, with just 12,000 adoptions from 150,000 customers (a rate of around 8%), Benioff said: “It’s the fastest-growing product in our history. There’s never been a faster-growing product.”
Benioff has also said that he would be the last Salesforce CEO who only manages humans. In a conversation with Axios, he said: “We are really moving into a world now of managing humans and agents together… Because I’m using Agentforce, I just have that much more productivity.”
Sounds good, right? Well, it might not be the best news if you work for Salesforce.
The Agents Don’t Work.. Yet
During his podcast with Dwarkesh, Karpathy said that we currently have some very impressive agents which he uses daily – naming Claude and Codex, in particular.
But he said there is still “so much work to be done”, adding: “We’ll be working with these things for a decade.”
When asked what the bottlenecks are, Karpathy replied: “Actually making it work.”
But, if he’s right in the apparent implication that agents are not yet properly working, then are AI-themed justifications for layoffs in the tech industry – by skeptics and advocates alike – a bunch of just-so stories?
We asked Senior Principal Architect AT at Aquiva Labs Robert Sösemann about whether he thought we were in the “year of agents”, the “decade of agents”, or neither – and whether AI agents really still are 10 years away.
He told Salesforce Ben he has great respect for Andrej Karpathy – and agrees with him.
Robert said: “At Aquiva, we’ve been building agents for about two years now – using OpenAI, Copilot, and Agentforce – so we’ve seen firsthand where things stand. There’s definitely been a lot of excitement, but in practice, not much has fundamentally changed yet.
“The models have improved, yes, but most agents are still unreliable. They don’t really learn, their memory is fragile, instruction-following isn’t dependable, and tool use – even with frameworks like MCP or agent-to-agent systems – is still in its infancy.
“You could call this the year agents ‘took off’ in terms of attention and experimentation. But for agents to truly replace classical software, we still have a long way to go. So I’d say Karpathy is right – this year is more like the spark of a hype cycle, and the real transformation will unfold over the next decade.”
Regarding Karpathy’s comments that the problem facing AI agent creators at the moment is “actually making them work”, Robert said: “I completely agree… Even within Salesforce, agentic systems haven’t really worked reliably yet. They’ve been too inconsistent for companies to trust them in areas where money or critical operations are involved. You simply can’t rely on them – or even test them properly in a consistent way.”
He added that Salesforce is now addressing some of these challenges with initiatives like the Agent Graph and Agent Builder, which aim to make things more robust.
But, Robert says, there are still many open issues to solve, including the ability of agents to learn by action and improve themselves over time.
“That’s something even a human intern can do, and yet today’s agents still struggle with it,” Robert said.
… Then Why Was I Laid Off?
In September, Salesforce Ben looked into whether AI was just an ‘excuse’ for layoffs across the tech industry, including at Salesforce itself. Benioff had earlier announced that the company had begun “rebalancing” its workforce by using AI agents to replace around 4,000 customer support division employees, sparking something of a backlash.
But, as Salesforce Ben journalist Sasha Semjonova explained in her article, this story is not unique to Salesforce – both Microsoft and SoftBank also made headlines with news of AI-focused layoffs this year – and some have been calling AI a “scapegoat” for the staff restructuring moves.
The 2025 Workforce Skills Forecast reported that eight million jobs in the United States alone would be impacted by agentic AI 2030, and that “workers who don’t adapt face a higher risk of role displacement.”
According to Layoffs.fyi, more than 112,000 tech sector employees have been let go in 2025 alone.
When asked whether he thought AI has been used as an ‘excuse’ for layoffs in the tech industry, Robert said that it is an oversimplification of the situation. He added: “Companies have always looked for ways to reduce costs and automate processes – that’s nothing new. What’s different now is that AI gives business leaders, for the first time, the genuine belief that this might actually be possible.
“That said, I don’t think we’re truly at the point where AI can replace people in a meaningful way. What we’re really seeing is a growing divide – between those who learn to use AI effectively and become 20-50% faster, more creative, and more engaged in their work, and those who reject it or dismiss it as a passing trend.”
AI is sometimes used as a convenient narrative for layoffs, but that’s only part of the story, Robert said. “There’s some truth in it – just not the whole truth,” he added.
Parallels to an Earlier Time
Are we in an AI bubble? The question is often asked – and, arguably, with good reason. If we look back to the dawn of the internet, we can see some overlaps with today: a huge amount of “hype”; discussions about how this new technology will change the world forever; talk of how early adopters will rake in the benefits – and the latecomers will suffer; and so on.
Referring back to this time might help clear our thinking up a little when it comes to the AI “bubble” – if that is actually what it is.
It’s undeniable at this point that the internet changed (and is still changing) the world, just like its advocates said it would. But the dotcom bubble burst all the same, and “pets.com” has become something of a punchline among people who know about this particular time in tech industry history. So… is that where we are now?
After asking Robert – who started studying computer science around the turn of the millennium, when all this was going on – what he thought of the comparison between agentic AI and the start of the internet, particularly the dotcom bubble, he said: “I definitely see parallels.
“There’s always an initial hype phase where the technology isn’t quite ready – just like with the early internet. Back then, we were still missing the fundamentals: reliable billing systems, proper e-commerce infrastructure, [and] integration between services.
“A few companies managed to build truly functional webshops, but most were just putting up flashy HTML and CSS pages, focusing more on web design than on usability or backend logic.”
He added that this trend seems apparent with every major technology shift – an early wave of excitement that overshoots reality, and then real progress begins.
“We’re seeing the same pattern with agentic AI today,” Robert said. “People are starting to grasp what’s possible and what the end goal looks like – but they’re ignoring all the missing pieces. That gap between ambition and readiness is what drives both the hype and the innovation happening in research labs right now.
“And it’s not just Salesforce – no one has truly solved it yet, not Lovable, not n8n, not anyone. Everyone’s essentially combining some instructions with some models and trying to make it work.
“The comparison with the dotcom era is absolutely valid. The only difference, I think, is speed – technology evolves much faster today, partly thanks to AI itself. So I don’t expect a major economic crash like the dotcom bust. Instead, I believe we’ll reach maturity much sooner – not in ten years, but in just a few.”
Final Thoughts: So What Does the Future Hold?
As Robert points out, we are seeing a pattern that repeats with every major technological shift. Just because there’s a hype phase and a cooling-off period, it doesn’t necessarily mean that AI is a bubble.
Just because agentic AI isn’t “ready yet”, as Karpathy put it, that doesn’t mean ‘Pandora’s box’ can be sealed again.
Companies that are successful today – perhaps we can name-drop one “Salesforce dotcom” – are the very same ones that started experimenting when the internet was not yet fully developed.
“They were building static sites and rudimentary online shops while others dismissed it as hype,” Robert says.
“It’s never about waiting for technology to be perfect. It’s about engaging early, understanding what’s missing, and building from there. That’s the only way to gain the experience and intuition that really pay off later.”
Have you been affected by tech industry layoffs? Email tips@salesforceben.com
