For the first time in history, Silicon Valley, the global heartbeat of innovation, is falling behind. Even engineering heavyweights and frontier labs are losing ground as users are demanding more than hyperscalers are prepared to deliver—a tension that’s reached a boiling point as teams like OpenAI race to ship both breakthrough capabilities and unprecedented uncertainty at the same time.
This is seismic on many levels. Historically, the best and brightest startups in Silicon Valley have been three steps ahead. They built businesses and legends around anticipating what people will want once the latest trend takes off. Google won the search war before the competition knew there was a war. Apple won the smartphone competition before Blackberry and Samsung realized they needed to accept that fate.
Today, Silicon Valley is moving too slowly to capture any titles, failing to keep up after pushing new social norms on everyone else for years. Worse, the biggest players are so focused on preserving walled garden business models, they aren’t innovating as big or as fast as they used to. And time is not on their side.
Social Demand Is Setting the Pace of Innovation
As we’re seeing with rapid model updates, sector-fluent plug-ins, and feature launches, pressure will only keep mounting on incumbents and hyperscalers to relent and give people what they want. The legacy playbooks of control and social guidance that ruled over us during the early days of the web, mobile, and cloud eras no longer apply.
The future of AI agents needs layers of coordination and orchestration that don’t care what kind of phone you have, what cloud provider your company uses, or which frontier lab’s models you prefer. They just work with everything.
That’s possible today, but there’s denial and determination among the AI leaders to keep users in one ecosystem. And it’s coming at the expense of the brand of innovation that put Silicon Valley on the map.
A solid example from recent days: progress now measured in days, not years: Anthropic chasing OpenAI with Claude for computer use. Even if they were planning to ship these capabilities, the announcement felt like a rush job in the wake of a watershed moment rather than a chess move from a position of strength.
Meanwhile, privacy and security have been thrown out the window by people believing they can customize an agent (or a swarm of them) with their massive backlog of emails, iMessages from friends, transcripts from deal meetings, and every other piece of software they have to deal with on a daily basis.
People Want AI Agents That Work. On Day One. With Any System.
Every hyperscaler and entrenched incumbent operating out of Silicon Valley is laser-focused on bringing AI tooling to every aspect of their ecosystem they can think of, regardless of customer sentiment or demand. And it’s not working.
People want, and increasingly expect, agents that can work like them, unconstrained by a single platform, product, or locked-down ecosystem. People want assistants for their personal and work lives that can do the same things they do. Everyone from Mark Zuckerberg to your neighbor is looking into building and deploying an agent that not only knows their job, but actively helps do it. It’s leaving both the frontier model companies and traditional software giants scrambling to catch up.
The promise of automating out the drudgery of work and home by hacking together armies of agents feels so tantalizingly close, yet just out of reach. The rewards seem to outweigh the risks, but no one is entirely sure.
The most jarring example of this has been in the heart of Silicon Valley itself. In a rush to prove to companies that agents can produce code just as well as engineers, early adopters are now drowning in an endless backlog of AI-generated code with no easy way to make sure all the newly written drafts are correct. The bottleneck has shifted from producing code to code reviews and even re-reviewing AI agent code reviews — a problem thrown into sharp relief by the AI agent-triggered AWS outage.
Engineers are actually spending more time double–checking what AI agents have written and waiting for backed-up testing pipelines to run. It’s swiftly getting to the point where everyone from the CEO to the newest backend engineer is questioning whether engineering teams are truly more productive.
If we put ourselves in their shoes, it’s easy to say that the next bottleneck is to make it easier to validate that code and unblock the engineers. And to look ahead even further, companies will then need to figure out how to have real-time testing agents double–check changes once they’re live to all users and automatically fix things before they break for everyone.
That might be a technically accurate assessment of where things are going with software, but it misses the bigger picture. We’ve already entered the era of innovation with non-technical people — not tech luminaries — in the driver’s seat.
What People Need From Their Agents Is, Simply Put, Agency
AI tools only go as far as their capabilities, which is why we see early adopters rabidly buying every Mac Mini in sight before throwing every piece of software they touch at them. Agents need to be able to draft messages, access email, review meeting decisions, track deal pipelines, understand brand guidelines, and so much more. No previously successful company can—or wants to—build that.
It’s an open secret that AI models are trending toward commoditization. Incumbents need to build for a multi-model, multi-everything future in order to close the innovation gap and retake a leading role in defining the agentic era of AI.
For example, NanoClaw was a herald for what agents should look like in the future. We’re already seeing this with the rapid adoption of tools like Town, for individuals who need a 24/7 assistant that is meaningfully productive. It’s hitting home at Crafting, where we’re seeing enterprise engineering teams wiring together data from a cornucopia of sources to build agents that can actually ship like software engineers. Even lower-level tools, like Zapier and Gumloop, are pulling in non-technical folks who are creating their own personal agent orchestration systems.
What this tells us: People aren’t opposed to AI if it can actually do the things we want. In fact, the things we want AI to do will bring the technology into mainstream daily life and into the enterprise at scale.
It’s not clear who will outlast the competition and emerge victorious when the agentic era hits its peak. But it is clear that, unlike every other innovation sycle in Silicon Valley’s history, startups will end up becoming the glue that ties everything together.
After decades dictating how the world experiences technology, the biggest players in town need to bend the knee to social demand and build agent ecosystems people actually want, or they will lose out entirely.
Sumeet Vaidya is the CEO and co-founder of Crafting, bringing enterprise-quality infrastructure to autonomous agents and engineers. He was previously an early engineering leader at Meta, Uber, and Discord.
