The agent boom is splitting the workforce in two

We’re only two months in, but 2026 is already shaping up to be the year of agents. The current surge began with Claude Code, which achieved critical mass over the holidays. That led to all kinds of lobster-themed software names (long story), which culminated in OpenClaw, an open-source agent creation and management system. It might also be a stealth marketing campaign for Apple to sell a ton of Mac Minis, but that’s neither here nor there.

It’s too early to say what kind of productivity gains the current wave of agents will create, but the push to agents is undeniable. It’s also very exclusive. For all the talk of, “the only coding language you need to know is English,” there are technical barriers to joining this wave. You don’t necessarily need to know how to code in order to use OpenClaw, but it helps considerably.

To help non-coders overcome some of the technical barriers to building and working with agents, AI companies have begun to release products that abstract away some of the more difficult parts. Anthropic released Claude Cowork—essentially Claude Code for the rest of us. More recently, Perplexity launched Computer, its “general-purpose digital worker” that users can prompt in natural language and watch it go to work.

It all sounds magical, and if you squint, you might even see a near future where knowledge work, and especially editorial work, transforms: instead of pulling levers on various software menus and dashboards, you’ll just talk to agents. They’ll handle the hard stuff, and if they run into barriers, you’ll just ask another agent to build the solution.

Where no code gets real

Back in reality, it’s not that simple. Even if you use one of the no-code systems like Claude Cowork, creating tools and workflows still involves breaking down processes, finding API keys, navigating permissions, and iterating continually. And even though Anthropic markets Claude Cowork for non-coders, when I used it, the app gave me instructions that included using the Terminal on my Mac—an app that most people have no idea exists. And if you don’t, you probably shouldn’t mess with it.

For builders, these don’t even qualify as barriers. Builders don’t need to be coders, but they do have characteristics that most workers don’t: They seek to understand the process beneath their tasks, and treat that process as modifiable and programmable. More importantly, they see failure and iteration as tolerable, even fun. They thrive in uncertainty.

But the reality is most people don’t think that way. We’ve trained a generation of office workers to work within software with clear boundaries and reusable templates. If there’s an issue, they call IT. Any feature request gets filtered and, if you’re lucky, put on a roadmap that pushes it out 6-12 months.

In short, most people don’t have a builder mentality to begin with, and expecting them to suddenly be comfortable working and creating with agents is unrealistic. In January, New York Times tech writer Kevin Roose pointed to a growing chasm between those fully in the AI bubble, who are building multi-agent teams to help them get work done, and those who aren’t, most of which have never even built a basic assistant like a Custom GPT or Gemini Gem. As someone who trains editorial teams on how to use AI, I can confirm this gap exists and is indeed massive.

All of this is to say, for all the hype you might see on X, the percentage of workers who have actually adopted agentic tools is extremely small. However, they’re still massively influential. The catch is that agents, at least as they exist today, are hard to deploy safely inside organizations. They need access to files, email, calendars, internal systems, sometimes the ability to take actions automatically. That’s not a tooling problem. That’s a permissions problem, and it makes security teams nervous for good reason.

You don’t need a sci-fi scenario to see why. A recent example involved an OpenClaw agent that appeared to run amok in a Meta engineer’s inbox, taking destructive actions despite attempts to stop it. Stories like that may be edge cases, but they point to a reality: delegating software access to agents can amplify ordinary mistakes into high-impact mistakes.

Clawing through permissions

Until there’s progress around security and fail-safes with these general-purpose AI agents, organizations will be slow to adopt them. That won’t keep builders—even those within those organizations—from developing or using them, however. They’ll just do it on their own time or elsewhere. This “capability chasm” between builders and users will eventually force solutions, and the systems those builders create will determine the workflows of the future.

For non-builders, this isn’t a great place to be. Becoming a builder, though easier than ever from a technical standpoint, means a shift in mindset that many simply aren’t up for. The alternative seems to be to sit passively, wait for agentic systems to filter down to you, and hope you don’t get laid off in the meantime.

There’s a third way, though. You don’t have to be a builder to understand how agentic workflows are changing your job. For journalists, that means identifying the parts of your work where human attention and judgment is paramount: the filtering of facts, the interviews, the writing (or maybe not), the cultivating of source and audience trust. You can help define what should never be delegated, and what can be automated without harming standards. You can push your organization—constructively—to adopt agents in bounded, defensible ways that match newsroom reality.

In other words, you don’t have to build agents to matter in an agent-driven workplace. But you do have to understand the systems being built around you, because soon enough, your job will be defined by defaults someone else designed. Most professionals will not build agents. But everyone will work inside systems builders create.