You should probably worry about your next job
Now might be a good time to find a place to spend the next few years
Last year was a very big year, I sold my AI startup and have moved to an IC (Individual Contributor) role at the acquirer. No longer being involved in the day to day management of a startup means I’ve had space and time to think clearly about what the future of work looks like and what the actual shape of future businesses will be. It also means I’ll likely be writing a lot more, which I am looking forward to. I’ve been kind of vagueposting on twitter about ‘you should probably worry about your next job’, so I decided to properly put the argument to paper.
The explosion of agents is going to radically change the way that we work. I use agents every day, to an intense degree, to do much more work than I could do by hand, a lot of it is work I’m glad I can delegate to a computer, but it’s work that is economically important. I’m a software engineer by trade and I haven’t written a manual line of code since September. This is a huge opportunity for those who are able to adapt, and for those who can’t, something worth preparing for. Things are about to get very weird.
A quick note on terminology. I’m going to use AI, agents, and tokens interchangeably throughout this. I mean an agentic system, one that can iterate and loop, browse the web, touch systems, send emails, read from APIs. A box of numbers that does a thing. If you put it in a harness (a software program where the LLM call runs in a loop) where it can interact with the world and tool call (talk to the web, talk to other companies, browse websites), it can do substantially more. A lot of this might seem far fetched now, but these systems are working in production in businesses today. It’s just a matter of time until the models improve or people build better harnesses and skills to enable them to handle more workloads.
Software is getting eaten by AI agents. This is not conjecture. Code is a perfect workload to hand to an agent because in compiled or typed languages with good testing, it’s easy to hand iteration to an agent to review and check its own work. The feedback loop is tight and the definition of success is clear.
I’ve spent a lot of time lately talking to other founders and people who run businesses, mostly technology but also lawyers, journalists, and people working in traditional engineering. What I’m seeing is newer businesses doing substantially more with way less, augmenting their workflows to an incredible degree with token usage. Marx was right that capital always has a desire to replace labour with machines. The difference now is you can replace human labour with compute at dramatically reduced cost, and without the coordination overhead that comes with people.
Businesses have a choice, work is delivered through either calories or tokens. You can spin up a new Claude or Codex instance, document a workflow in a few minutes, plug it into your systems, and get decent output in a few hours. Hiring is an intensive and fragile process. If you’re a business owner with deadlines bearing down, the choice is clear. Tokens don’t take holidays, get sick, or quit. Tokens also crucially can be thrown at the same boring problem over and over again without getting burned out or frustrated. There are some jobs that are perfectly suited to computers.
When there’s a business imperative to have something produce economically valuable output, businesses are going to spend time making an AI system work at that problem to 70 or 80 percent of human quality. Particularly when the alternative is to expand headcount. The comparative downsides of labour versus AI are just too stark.
This is exactly what Acemoglu has been warning about: pure substitutability of work and workflows to AI, without the creation of second order, better jobs for the people being displaced. The difference between Acemoglu’s papers and what I’m describing is that I’m watching it happen in real time, from insights into businesses doing it. This isn’t modelling anymore.
We already have entire workloads for which AI is perfectly substitutable at 90 percent or above. Copywriting, proofreading, and copy editing are jobs that AI does an incredible job of right now, helping millions of people produce economically valuable output. The previous work of subeditors, editors, and proofreaders is being ablated away. These systems are now turning on BPO and people doing labour arbitrage in the global south. It’s easier to reach for a token system than it is to reach for Upwork. This will only accelerate. Anthropic is directly partnering with finance orgs to integrate AI to replace and accelerate workflows. The model providers have said this is the work they want to go after.
It’s happening at pace. The tools are good and getting better. Excluding building software, I’ve put together financial projections and unit economic evaluations using Claude to generate Google Sheets, I’ve used Codex in a harness to drastically reduce the total bill of an AWS account, I’ve had Claude research and evaluate prospects prior to reaching out and figure out who best to target. These models are incredibly capable and we’re just scratching the surface of what we can do.
These are types of work that agentic systems can evaluate and provably solve because they have specific inputs and specific outputs. That makes them straightforward to optimise. The other, important but secondary piece is skills, which are basically markdown files that tell an agent how to do something. They range from SEO optimisation to managing Meta ads to using CLIs and APIs to do work. You are now seeing open source software ship as Agent first, shipping skill docs alongside their actual code and docs. Agents perform these workloads better when the substance of the work is effectively documented, and they can pursue those workloads again and again, refining the skill document so output continues to increase in quality and consistency.
An interesting side effect of substituting humans for token systems is a drastic reduction in coordination costs. Coordination will switch from human boundaries to API boundaries or machine boundaries. Even existing systems of record effectively become an API with agents using browsers. These agent systems communicate by writing things down in exhaustive detail (if you set them up correctly).
Systems of record become more important because they give the agent context about what worked, what didn’t, and what to do next - those systems of record become crucial, as a canonical record of actions inside your organisation. Coordination becomes cheaper, faster, and less fragile. Agents don’t need meetings, they just need context. There will obviously be resistance to change in organisations, particularly by those who benefit from the current structures of access to information. That resistance will itself incur downstream costs in terms of margin, ability to react to market demands and organisational pace itself.
The crux of the issue is whether the work is done to an acceptable standard. That’s always going to be a matter of management and taste, but crucially it comes down to whether the output is economically valuable or not. For another essay but cheap tokens make a huge amount of work economically viable that wasn’t before. The calculus has already shifted, and these tokens are likely the most expensive they’re going to be for this quality of output.
AI sceptics will say these systems are producing a facsimile of work, that the facade collapses on close inspection. There’s something to that but through iteration and development you can drastically improve the output of these systems. At some point an effective enough facsimile that achieves the same aims is substitutable for the real thing. That point is closer than most people think. People are already paying good money for the output of LLMs.
So why do I say you should worry about your next job? We’ve been hearing about doom and gloom regarding labour substitutability for years, and apart from a handful of cases there haven’t been true mass layoffs with staff being replaced with AI.
My bet is there won’t be mass layoffs in a dramatic sense either. What we’ll see in existing organisations is headcount will not be replaced after attrition. As people leave there’ll be an incentive to replace those workflows with agentic systems. If you’re already using these systems you’ll be able to do more with less, so hiring gets constrained. There’s an economic and managerial incentive to do this. Managing people is stressful and messy and wonderful, but managing robots is substantially easier. If there are hires they will be in very specific roles, either managing these workflows, creating new ones, or doing work AI cannot do in the white collar context (which I’ll outline below).
Larger organisations are already moving. The smart ones are doing the AWS approach, opening up internal data silos and exposing authenticated APIs that agents can touch. If you write software day to day, the lines of code you write are going to decrease, maybe the amount of software you review increases, but that will likely change as well. This has already started and this year is the inflection point. Any work that can be proven (as in has a provable, verifiable answer), like software or finance work, will accrete more and more toward agentic systems doing the drudge work.
Where AI is really going to land like a rocket is on people whose entire job is coordination and sharing of information. If your job is to build decks, get people together, share information upstream or downstream, that work is going to get completely delegated to agent systems in the next 24 months in any savvy large organisation. CFOs are looking at time costs in meetings with barely concealed glee and are desperate to pare them down. All of this can be executed on right now with technology we currently have. The models will almost certainly get better this year, but even if they didn’t, this work can already be mostly automated. It’s just a matter of corporate will, the incentives for businesses to do this are incredibly clear. Reducing coordination costs increases speed of execution. Previous coordination roles will start to look more like product management, but the product that is being managed is the output of the business division.
There’s obviously work agents won’t replace. Proper GTM and outreach, real sales, a lot of work in the physical world as well as areas with significant regulatory burden (I’ll write about the durability of regulated organisations another day). But in the white collar world where you produce designs, legal briefs, software, articles or any number of digital goods, the number of roles needed to produce the same economic output is going to keep decreasing.
For founders the opportunity is obvious: stand up a lean team (maybe even just yourself!), focus on revenue and retention, let agents handle everything else, and watch your revenue per employee compound. Every hire is an attack on margin. The businesses that internalise this earliest will have structural cost advantages that are very hard to close allowing them to take on larger incumbents who incur the costs twice, first as pure headcount then as coordination.
For workers the question is harder. If you really dislike your current job, you should look for a new one, quickly, because you might be there for a while. The answer also isn’t to avoid the tools, it’s to become the person who knows how to drive them. The people who will be fine are the ones who can evaluate output, direct systems effectively, and tell the difference between something that’s economically valuable and something that isn’t. That’s a management skill as much as a technical one. Weirdly, being an Engineering Manager for 15+ years put me in a very good position to know how to use AI agents very effectively.
It’s going to be a weird couple of years.