We're now well into 2026, and the dust is starting to settle on a year that many—at least in certain circles—had confidently proclaimed would be "the year of the agent." Remember that? The promises were grand: AI agents transforming workflows, replacing tasks, and ushering in a new era of autonomous systems. But did it actually happen?
That’s precisely the question Ryan Donovan, host of the Stack Overflow Podcast, recently put to Stefan Weitz, CEO and co-founder of the HumanX Conference. The pair sat down to dissect AI’s real-world evolution over the past year, digging into why the agentic future hasn't quite manifested as prophesied, the industry's quiet retreat from the grand ambitions of AGI, and the persistent roadblocks hindering broader AI adoption—everything from an inherent distrust of unpredictable systems to the nitty-gritty of enterprise data readiness.
Credit: Alexandra Francis
## The Agent Reality Check
Stefan Weitz doesn't pull any punches when reflecting on 2025's "year of the agent." At last year's HumanX Conference, we heard plenty of speakers touting the imminent arrival of agents everywhere. Yet, as he sees it, 2025's reality was a stark contrast to the promised "utopia." It’s a familiar pattern in AI: a generous dose of hyperbole, perhaps fueled by genuine belief in a product's potential, often outstrips short-term capabilities. We’re overestimating what can be done in the short term, while underestimating the long-term potential—a sentiment that echoes a well-known, if sometimes misquoted, observation from Bill Gates.
What we're seeing now, thankfully, is a shift. The fever pitch of "agents will replace everyone's jobs tomorrow" has cooled. We're moving into a more pragmatic phase where we recognize that these systems certainly make sense for "certain disciplines," especially in coding. But this pragmatic view also highlights the tough practical challenges. Managing these agents is anything but simple. Building multi-node agent architectures is a headache, and most existing infrastructure simply isn't ready for agents that might need to run for hours and then autonomously take action. The aperture on the agent ecosystem is opening, revealing the real work needed to make them truly functional and useful beyond just being "cool."
## The Core Obstacles to AI Adoption
So, if agents aren't quite the panacea we were sold, what's holding them back? Stefan points to three significant gaps in the current technology stack:
### Infrastructure Isn't Ready
First, there's the underlying infrastructure. AI-ready data centers with advanced networking? They're largely missing, particularly in older, legacy organizations. We also lack robust multi-node architectures that allow agents to work in concert effectively. And don't even get started on seamless support across disparate cloud environments and edge locations—it's just not there yet. These are fundamental hurdles that demand solutions before agents can truly scale.
### The Trust Deficit
Then there's the question of trust. This one feels like a glaring contradiction: over 80% of developers plan on incorporating AI into their coding projects this year. Yet, almost half of all developers—a number that's doubled recently, according to what sounds like a Stack Overflow DEB survey—simply don't trust it. That's a significant dichotomy. It's not hard to see why: AI models are inherently non-deterministic, meaning their behavior can be unpredictable. When you throw in multi-cloud and multi-edge environments, you're introducing a whole new layer of vulnerabilities. Without a foundational layer of trust, broad adoption will inevitably stall.
### Data, Data, Data
Finally, and perhaps most critically, machines still need machine-readable data. Here’s the thing: most enterprise data isn't set up for consumption by an agent. It’s trapped in old ETL pipelines, buried in batch jobs, scattered across flat files, or worse, residing on an AS400 mainframe somewhere in a corporate basement. This data needs significant transformation to become accessible to an agent. While new systems claim to read unstructured data in clever ways, Stefan rightly emphasizes that clean, well-structured data will *always* outperform a machine attempting to make sense of a mess. Organizations are jumping into an agentic future without truly grasping that their data isn't prepared for it.
## The Retreat from AGI and the Rise of Practical AI
One of the most telling shifts over the past year (2025-2026) has been the quiet fading of the AGI discussion. That grand vision of artificial general intelligence, the "golden fleece" out in the distance, has largely become passé. While it remains an incredible long-term goal, the industry's focus has sharpened on immediate, "efficacious applications" within narrowly scoped vertical areas. Think customer service, legal tech, and healthcare. That's where AI is truly delivering value right now.
This is a pragmatic concession, and frankly, a welcome one. Even at its most basic, generative AI offers incredible utility as a natural language interface or for semantic search. It doesn't need to be a PhD candidate to be useful; the foundational advancements in human-computer interaction models are significant. Stefan, for his part, sees these models as "stochastic parrots"—powerful pattern-matchers, but not, by definition, a new form of intelligence.
## AI and the Evolving Developer Landscape
The impact on developers is profound. LLMs are unlocking "developer adjacent" capabilities for people who, like Stefan, haven't written production code in years. They lower the barrier for syntax recall, letting people focus on core principles like security and database structures. You can spot a bad data model recommended by an AI if you understand the fundamentals.
And yet, this power comes with a caveat. "Vibe coded" applications—built quickly with AI without deep engineering oversight—often look great on the surface but quickly fall apart in production. They don't scale, suffer from terrible data models, chew through CPU cycles, and are essentially "automated tech debt at scale." Tools like Claude Code might inadvertently lead to some "really bad decisions" if not guided by sound architectural principles.
This highlights a critical point: while the "programmer" part of development might become less about rote syntax, the "engineer" part is more important than ever. Stefan illustrates this perfectly with a recent anecdote about troubleshooting his wife's locked Mac. ChatGPT, when asked for help, made a statistically likely assumption (forgotten password) and failed to suggest a specific solution that only became apparent after deeper investigation. The AI didn't "know" she actually had the right password but was using a Spanish keyboard with an inverted exclamation point. It picked the most probable scenario.
The takeaway for developers is clear: precise prompting is paramount. Tools that help define a target via robust markdown files for specs and requirements are becoming increasingly valuable. The better you define the problem, the better the agent’s output. We’re not yet at a point where an AI can truly be your architect, making complex design decisions from vague inputs. The human still needs to understand the principles and guide the machine effectively.
If you're interested in connecting with Stefan or exploring the future of AI further, you can find him on LinkedIn. The next HumanX Conference is set for April 6-9 in San Francisco, building on the discussions we heard in last year’s episodes recorded on the conference floor.
*(And a quick shout-out to humblebee for earning the Populist badge with their excellent answer to "How to open/run YML compose file?")*Here's the thing: everyone's talking about AI, but the sheer volume of capital pouring into the sector still feels, well, *irrational*. Stefan Weitz, CEO and Co-founder of HumanX, put it bluntly: "We're in an exceptionally frothy period." And he's not alone in that assessment.
The AI Investment Treadmill
Weitz, whose venture fund sees pitches constantly, notes that what a "good team and good idea" fetches in valuation today is ten times what it would have just two years ago. The underlying issue, he suggests, isn't just plain old FOMO (though that's certainly a factor). It's a deeper uncertainty among investors. "People still don't know what the ultimate potential of this technology is, 'cause we're all still figuring it out." That nervousness drives some "somewhat irrational" bets, hoping that *one* will hit big.
The challenge for many venture capitalists isn't simply evaluating a standard SaaS product anymore. Those, he points out, were largely comprehensible: 'I get how this works, I get the backend technology.' With many AI products, you often need a PhD in Computer Science to even grasp the pitch. While some highly technical firms like Lux Capital exist, a significant portion of successful funds simply don't have that depth. So, instead of deep technical diligence, they lean on founder track records – 'She's good, she's done three companies, she was at Google for 10 years, I'll trust her.' It's a high-stakes bet on the jockey, not always a clear understanding of the horse.
Giants and Goliaths: A Market in Flux
This influx of capital isn't just inflating seed rounds; it’s fueling the "massive capital requirements" of central AI players. Weitz draws a parallel to the late 1990s fiber market: over-deployment, too many cables, a glut that wiped out many players. He expects a similar dynamic in AI, even for the big names, forecasting changing fortunes rather than outright collapse.
We're already seeing shifts. Take the competitive landscape between ChatGPT and Gemini. Just twelve months ago, ChatGPT commanded an 86.7% market share, with Gemini trailing at 5.7%. Today, ChatGPT holds 64.5%, while Gemini has surged to 21.5%. Much of Gemini's growth, it seems, has come directly at ChatGPT's expense. This isn't necessarily a judgment on quality, but a stark reminder of the institutional advantage held by giants like Google and Meta, whose balance sheets can sustain "incredible investments."
Google, in particular, looks set to dominate. Weitz, who worked on Bing's search engine years ago, recounts how query lengths have exploded from a mere 2.4 words to "two-paragraph queries" that Google's AI mode now brilliantly handles. With its "surface area, technology, user base, [and] trust," he believes it's "Google's to lose." They have the components, and Gemini, he admits, is "really impressive these days." The big question isn't *if* someone can beat Google, but *who* from the current crop might manage it.
The Art of Prompts and the Singularity Myth
Amidst the market froth, there’s also the question of AI’s actual capabilities. When the discussion turned to Ray Kurzweil's reported 90% AI-generated code commits to "Claude Code," the 'Wish.com Singularity' jab was pretty sharp. Weitz dismissed the idea of genuine AI autonomy, instead comparing it to von Neumann machines – self-replicating, yes, but still a distant future. He attributes Kurzweil's impressive output to "crazy good prompts" and a deep understanding of architectural principles, not the AI working unguided.
"You gimme a canvas and paint and I can crank out a picture," Weitz quipped, "It'll look terrible." The analogy holds: true artistry still requires a human hand, even with powerful tools. This isn't AI spontaneously engineering software; it's skilled engineers using AI as an extension of their expertise. The notion of simply telling an AI, 'Hey, crank out version 1.3, figure out the features,' would, as Weitz put it, "go off the rails so fast." It's still, fundamentally, engineering.
The Road Ahead: Agentic AI and Old Problems
The future of AI is exciting, with "agentic AI" at the forefront. Weitz highlights several critical areas for development: better "memory architecture" (including "forgetting and compacting"), "multi-session memory," and vastly improved "evaluation benchmarks" beyond simple quiz-taking metrics like MMLU and Helm. The challenge of "error accumulation" in complex multi-chain agent systems is another pressing issue.
Interestingly, Weitz finds himself revisiting old problems. Multi-Chain Protocols (MCPs), for instance, remind him of Microsoft's Common Object Model (COM) from the 90s, complete with "DLL Hell," registry issues, and security oversights. "Security is an afterthought" in MCPs today, just as it was in COM. "Nothing is new anymore," he sighs, "We're literally revisiting the same lessons from 30 years ago." As Ryan Donovan succinctly put it, agentic models are "speed running the service-oriented architecture pipeline."
HumanX: Bringing Focus to the Conversation
Given all these complexities, hype, and technical challenges, the HumanX Conference, founded by Weitz, aims to cut through the noise. Set to take place in San Francisco, it boasts an "unreal lineup," including Vint Cerf and Al Gore, alongside industry leaders like Brett Taylor, Jaime Teevan, Fei-Fei Li, and Matt Garman.
What makes HumanX stand out? No one can pay to be on stage. "Everyone here is truly editorially chosen," Weitz stresses, promising "exceptional insights you can't get by just reading the news." Crucially, the conference focuses on job roles. The entire agenda is tagged to common enterprise functions—marketing, technology (dev and IT), sales, operations—allowing attendees to filter sessions specifically relevant to their work.
Beyond structured content, the conference emphasizes interaction. Over a third of its 300+ speakers will lead "small group sessions" and "chalk talks." Weitz recalls even industry titans like Vinod Khosla engaging directly with attendees last year. The goal is clear: "You leave this thing with practical guidance, with partners you can trust, with vendors you can trust, with connections you can call when you get stuck, with knowledge you couldn't get from reading the web."
Looking Forward
The conversation around AI remains in its "figuring out" phase. Weitz is particularly keen on the progression of agentic discussions, exploring memory architecture and robust evaluation metrics. He's also excited about "physical AI" and robotics, despite joking about robots murdering him in his sleep. His candid observations on the AI bubble, the market's evolving dynamics, and the constant re-learning of old lessons underscore the urgent need for a more grounded approach.
Ultimately, while the transformative potential of AI is immense – in areas like ag tech, drug discovery, and small molecule discovery – the industry needs to dial down the "hype and hyperbole" and embrace "truth in advertising." Events like HumanX are a critical step in fostering that necessary conversation, moving beyond the noise to genuinely understand what AI *is* and what it *can actually do*, rather than what we merely wish it could. The journey from frothy valuations to truly impactful applications will be long and complex, but clarity and connection will light the way.
**For more details on the HumanX Conference, visit [humanx.co](humanx.co).**
**You can connect with Stefan Weitz on [LinkedIn](https://www.linkedin.com/in/stefanweitz).**
**Ryan Donovan, host of the Stack Overflow podcast, can be reached at [email protected] or found on [LinkedIn](https://www.linkedin.com/in/ryandonovan).**
