We’ve only got a couple of months left in 2025 that was supposed to be the year of AI agents. I think at this point we’ve got sufficient data to evaluate whether this prophecy was met or not. Did agents go beyond a buzzword and turn into something that is changing how we work?
To recap, back in January I said that I don’t quite believe the claims made by Charles Lamanna. His prediction was that by the end of 2025 we’d all have a team of agents working for us. At the same time, outside the tech vendors, AI enthusiast Simon Willison made his own predictions about what would happen in AI in 2025. He was expecting people like Charles to experience a disappointment in regards to the agentic dreams becoming a reality.
2025: the year of agents. Predictions from Charles Lamanna and Simon Willison.
Recently, Simon admitted that according to the definition he has been using, 2025 had actually turned into the year of AI agents after all. Now, we must keep in mind that for software developers that are extremely deep in this tech (Simon coined the term “prompt injection”, for instance) the future may already be here. For us mere mortals who are not writing our own software but rather using and implementing tools from Microsoft, our reality looks quite different.
Still, I do think it’s encouraging that notable figures in the AI community who are known for their balanced takes on tech, rather than breathlessly spreading hype, are saying agents deliver value to them already. I am not fundamentally against the idea of GenAI technology maturing into useful everyday tools. Even with the limitations of LLM, given the right level of expectations and sufficient guardrails, they could do more good than harm. One day.
Personally, I think the word of the year in 2025 should not be “agent” but rather “vibe”. In February, Andrej Karpathy made the term famous with this post of his:
“There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”
Talk about how simply giving a name to something can launch a movement! Corporations like Microsoft only wish they could achieve an impact like this with their Copilot marketing campaigns. Yet half a year later, we saw how MS was now latching themselves onto the concept Andrej made famous, with the “vibe working” message used to introduce Agent Mode and Office Agent in M365 Copilot.
In 2025, everyone’s been vibing with AI. As for agentic AI — we see little traces of it outside vendor and vendor partner marketing pitches. The one arena where some level of agent adoption seems to be happening among actual end users who are not directly selling them for a living is coding. Simon Willison is an example of that. Yet so is Andrej Karpathy who co-founded and worked at OpenAI, attempted to make Tesla’s autopilot happen for a couple of years, and more recently started building the AI education company Eureka Labs.
He recently appeared in a podcast and talked about, among other things, the usefulness of AI agents in the software he’s been building. Such as nanochat, “The best ChatGPT that $100 can buy”. According to Andrej, in such projects this agent-led coding was not useful because the project was "intellectually intense code" where "everything has to be very precisely arranged". He described the generated code as slop - something that could well have been the “thing of the year” in 2024.
As an example of what useful LLM applications in the year 2025 look like, I didn’t actually watch that 2.5h video at all. I’m sure it’s great, yet on a Saturday morning when the kid woke us up at 6 AM and I had only a few minutes here and there - spending a blockbuster movie’s worth of time simply wasn’t possible. But AI made it possible for me to consume the key points of Andrej’s talk in 15 minutes. Via a mind map that Google’s NotebookLM generated:
LLMs can be seen as calculator words, and they can do a beautiful job with words in the right context, with the right expectations, and the responsible application of the technology. In contrast, LLMs in 2025 make for a terrible AI agent, and Andrej agrees with me. He says that we should be talking about the decade of agents instead:
Karpathy stated that he was "triggered" by the "year of agents" idea because he felt there was "some over-prediction going on in the industry" regarding timelines. He believes the situation is "more accurately described as the decade of agents", explaining that his reaction implies "we'll be working with these things for a decade" because there is still "so much work to be done". He predicts it will take about a decade to work through critical issues such as lack of intelligence, insufficient multimodality, and the inability of current models to handle computer use and continual learning, which prevents them from functioning effectively as an "employee or an intern".
Rather than me just citing LLM outputs to you for the rest of this newsletter issue, I’ll instead reflect on how the words from a top AI expert align with the thoughts that a low-code architect from the Microsoft BizApps domain has formed over the year 2025.
Subscribe to Plus to read the rest.
Become a paying subscriber of Perspectives Plus to get access to this post and other subscriber-only content.
Upgrade