The AI Paradox: More Automation, More Humans, More Work
My previous prediction about Claude Code for non-engineering work was, to my pleasant surprise, unbelievably right. A year ago, I observed that people were "sleeping on Claude Code" for non-technical tasks like fixing files or sorting hard drives. This insight has since played out dramatically, with the emergence of tools like Cohere and Codex leaning into non-technical uses of coding agents. This unique perspective comes from living in the future of work at my company, Every.
At Every, we're almost 30 people strong now, having doubled in size in the last year. Despite being an "AI-forward" startup where everyone, from engineers and designers to writers, editors, salespeople, and customer service, is an AI early adopter, we actively hire more humans. Our approach isn't to prognosticate the future but to "just live in it together." This environment, coupled with early access to alpha and beta models, allows us to notice and articulate emerging trends. Our experience with Claude Code was organic: we tested new models, and when Sonnet 3.5 or 3.7 arrived, we realized its power for general work. We then saw people organically "reach for it" – what we call the "reach test" – which confirmed its utility beyond niche technical applications.
I have three main predictions for how the way we work is going to change, the shape of the work we'll be doing, and who will be most successful in this coming future.
Prediction 1: How the Way We Work Will Change – The Bifurcation of Agent Interaction
My core prediction for how work will change is that it's going to bifurcate into two main ways regarding how we use agents. We will look back in a year and realize we actually have more work to do, not less, even as models get better at automating tasks.
Delegating to Super Agents
First, everyone in a company will have at least one agent that they talk to and can offload work to. Initially, I was convinced that personal agents—like having a "daemon" on your shoulder as in The Golden Compass—would be the primary model. However, I've completely flipped on this.
"In order for an AI agent to be useful right now, it really needs a human who cares about it."
What we've found is that personal agents often break and require too much maintenance for most people to bother with. Instead, the model that is working is a "super agent, like one agent for the entire company." We're already seeing this at companies like Shopify with their agent River, and Ramp. This model typically involves a forward-deployed engineer or someone with a similar profile being responsible for maintaining and ensuring the agent is useful for the whole company. I believe this will start at the top, with general company-wide agents, and then trickle down to more specialized team agents as models become more independent and less "fiddly." For work, these agents will likely live in Slack; personal agents will handle "computer errands" like ordering groceries.
Codex or Co-work as Your Operating System
The second, and perhaps more profound, shift is that most of the work you do will actually happen on your computer in an environment like Codex or Cloud Co-work. These tools will become the operating system for all your work, from email to document creation.
Anthropic first realized the power of putting a coding agent directly on your computer. With access to the terminal and everything you have access to, these agents became incredibly powerful for any kind of work. While OpenAI initially lagged, their recent Codex Desktop app has, in my opinion, surpassed Anthropic by skipping straight to the ultimate paradigm.
Codex is now my daily driver. When I'm writing a document, I open it in a Codex thread, which has an in-app browser. I use it to access Proof, our online markdown editor. Codex runs alongside me, watching what I'm doing and having full access to my computer.
"I basically feel like I have this parallel work buddy that not only can it like respond and write in the document, but then it can go do research. It can go it can use my computer to basically do anything that I can do on my computer."
This has been transformative. For example, I've maintained "inbox zero" for 10 straight days by having Codex, with our email agent Cora, gather all my emails, render them on a page, and then I simply monologue instructions. It can research, gather documents, and handle tasks I'd normally procrastinate on.
The critical insight here is a reversal: instead of embedding AI into a browser, we're putting a browser inside the AI agent. This "magical combination" allows the agent to see everything you're doing.
The SaaS Paradigm Shift
This shift has a significant second-order effect: SaaS tools will run within Codex or Cloud Code.
"I would buy SaaS stocks right now. I would I think the SAS apocalypse is done and SAS stocks will be up majorly in the next couple years. Not Not investment advice, but you know, I would buy SAS stocks."
When I use Proof or any website inside my agent, I'm using my tokens, not the vendor's. This fundamentally changes the business model for SaaS companies, potentially saving their margins. They don't have to bear the token costs. What agents actually do is increase the number of users of SaaS, not get rid of it, leading to an "insane spike in demand."
Companies like Cursor are moving in a similar direction, realizing the importance of a "harness" for models, though Cursor has distinctly chosen to focus on programmers. For other product builders, preparing for this future means designing software for both humans and agents to collaborate seamlessly.
- Simpler Products: You can make products simpler because agents handle complex formatting, tables, etc.
- New Affordances: Software will need approval flows, summary inboxes, detailed logs, and quick rollback capabilities for agent actions, as agents can perform a "billion different things" to a document simultaneously.
- Infrastructure Challenges: Products must handle agents making "a billion requests in like 3 seconds."
- Agent-Sent Bug Reports: Imagine an agent autonomously sending a detailed bug report with exact repro steps and even code suggestions to a company's agent, creating a "very fast like closed loop."
This paradigm also signals the end of the "CLI era."
"CLIs are over. We speed ran the CLI uh era. It was nice while it lasted, but I think it's pretty it's pretty clear It's not that CLIs are going to completely go away... but it's more or less they're using Codex, cloud code, cursor, um that kind of thing."
GUIs are simply nicer, and they can provide all the same benefits, especially for non-programmer work. Most technical people at Every are no longer using CLIs as their main work surface.
Finally, "two agents are better than one." My Codex can interact with another agent, providing far more context about me and my intentions than I could ever type, leading to a significant speed-up. If I'm setting up a new agent product, instead of a manual onboarding checklist, my Codex can simply provide all relevant context about my work, my preferences, and even fix issues if something goes wrong.
Prediction 2: The Shape of Work Will Change – More Human Work
Despite the incredible progress in AI, my team at Every doubled in size last year. This highlights a crucial point: "Automation is a lie." Every automated task still requires a human to oversee it and ensure it's working effectively.
"Every time you automate something, in order to make sure the automation is working well, you need a human on top of it like making sure that it's working well."
I call this the "allocation economy," where humans act as managers for AI, constantly checking in, refining, and guiding the work. The discourse around AI often misses this, partly because benchmarks can be misleading, making AI seem more autonomous than it truly is.
I created a "senior engineer benchmark" based on an app I "vibe coded"—Proof. After its launch, it was plagued with issues that I couldn't fix. I hired two senior engineers to rewrite the codebase independently.
Then, I challenged new AI models with this prompt:
This is vibe coded slop. If you wanted to rewrite it from first principles, how would you write it? Go do it.
Until GPT 5.5, models scored around 30 out of 100 (human senior engineers scored in the high 80s, low 90s). GPT 5.5, using an Opus 4.7 plan, achieved a 62. Crucially, GPT 5.5 was the only model that had the "sense of agency and confidence to just like rip out old code and just like actually rewrite from first principles." Other models would "paper over the edges" despite explicit instructions not to.
The key difference between humans and AI is the "higher frame." If I asked a model to fix a list of issues, it would diligently fix those issues. A human senior engineer, however, would look at the codebase and realize, "This is a piece of [__]... We're going to have to like actually rewrite a lot of this." Models don't typically do this on their own.
Benchmarks measure problems we frame and can articulate. The human act of thinking, framing the problem, and prompting the model is the unmeasurable, irreplaceable work. This is why even with exponential model progress, I continue to hire engineers.
It's also interesting to note that human-written code from before 2021-2022 is now considered very valuable data for training these models. This "original human code" is a testament to the unique thought process. While my benchmark allowed engineers to use AI, they leveraged it in a way I, as a non-expert in that codebase, could not. It's a race between humans using AI, because "AI doesn't use itself."
Prediction 3: Who Will Be Most Successful?
In this future, "product managers and designers are going to do very well."
"Creativity. It just feels like it's going to be more and more valuable to stand out from all the slop that people are shipping and launching constantly."
Models make "yesterday's human competence cheap" by commoditizing it. The human role is to take this "frozen human competence" and use it to "make something new and interesting." The ability to frame problems at a higher level, to understand user needs deeply, and to design novel solutions will be paramount.
Actionable Takeaways:
- Embrace Agent-First Workflows: Start experimenting with super agents for company-wide tasks and desktop agent environments like Codex or Cloud Co-work for personal productivity. Get comfortable with agents as your primary work surface.
- SaaS Companies, Adapt Your Products: Prepare for a future where users bring their own AI agents into your software. Design for seamless collaboration between humans and agents, focusing on simplified interfaces, robust backend infrastructure, and features like approval flows, logging, and rollback for agent actions.
- Invest in Human Oversight: Recognize that automation is not full autonomy. AI agents require human care, maintenance, and strategic direction. Companies will need to staff roles dedicated to "managing" these agents, refining their performance, and framing the higher-level problems for them to solve.
- Cultivate Creativity and Framing Skills: As AI commoditizes existing competencies, the most valuable human skills will be creativity, critical thinking, and the ability to define new problems and solutions. Product managers and designers, who excel at framing challenges and envisioning novel experiences, will find themselves uniquely positioned for success.