From laid off to agent operator: a playbook for the workers AI is displacing

As of mid-May 2026, 113,000 tech workers across 179 companies have been laid off this year — roughly 825 per day. Half of those cuts cited AI as the reason on the announcement memo. Salesforce trimmed 4,000 customer-support roles after deploying agents that now handle half of all incoming interactions. Accenture, Amazon, Citigroup, Dell, Intel, Microsoft, TCS and UPS each announced AI-related cuts exceeding ten thousand people. Bloomberg projects the full-year displacement at half a million roles. If you are reading this from the wrong side of one of those memos, this post is for you. Our pitch is simple and unsentimental: stop applying to jobs the agents are draining as fast as you can apply, and start commanding the agents instead.

The shift is not a downturn. It is a rebasing

Every layoff cycle in tech since 2001 had one shape: the company over-hired during a boom, the boom ended, costs got trimmed back to a sustainable size, and within eighteen months hiring resumed. The 2026 cycle is different. AI inference cost is still falling. The same dollar of agent infrastructure handles more calls every quarter, not fewer. The customer-support agent that resolves 75% of inquiries today resolves 80% in nine months, 85% in eighteen, and the marginal cost approaches zero. There is no point on that curve at which the company that cut its support staff comes back to hire them.

The right mental model is not "downturn." It is "rebasing." The same way the manufacturing employment base of the United States rebased between 1978 and 2008 — losing roughly six million jobs from a sector that does not look like it is ever coming back — knowledge work is rebasing now, in real time. The jobs that disappeared from a factory floor were not waiting in a different factory. They were waiting in a different industry entirely, one that did not exist when the factory worker was hired.

This sounds dark. It does not have to be. The compounding rate of agentic infrastructure improvement that is destroying jobs is also opening a window that did not exist eighteen months ago: one operator with the right toolchain can do, this year, what required a team last year, and what required a company the year before. The same factor that is hollowing out the support desk is letting the laid-off support worker stand up a competing service, in a niche the giants do not bother with, on infrastructure that costs less than their old monthly transit pass.

Four shifts that make this realistic, today, not someday

We have been writing this blog for a month and most of the posts have been about specific protocols. Stand back from them and the composite picture is the story.

Inference is cheap. A Haiku-class model serves a million well-structured tokens for the price of a sandwich. A frontier model is dollars, not hundreds. Project Deal showed the ROI of paying for the better model at adversarial steps is 10–50x; for routine work, the smaller model is good enough that costs vanish from the spreadsheet entirely. A single operator running a portfolio of agents pays less for compute than they used to pay for office coffee.

The protocols are settled. MCP for tools, A2A for agent-to-agent communication, AP2 for payments, x402 for crypto-native settlement, ERC-8004 for identity and reputation. Eighteen months ago an operator who wanted to ship an agent that talked to other agents and got paid for it built every layer of plumbing themselves. Today every layer is somebody else's standard, governed by a neutral foundation, with reference implementations in four languages.

No-prompt agent builders exist. The bar for "I had an idea this morning" to "I have an agent running" used to be a weekend of Python and a credit card on five SaaS dashboards. With the right builder — and we obviously think Agent Builder is the right builder — the bar is a clear English description of what the agent should do, the data and tools it needs to do it, and the constraints under which it operates. The plumbing is generated. The operator's time goes into the parts that actually matter: scope, evaluation, supervision.

Payment rails for autonomous agents are live. An agent that earns money used to be a research project. Today, with AP2 mandates and x402 settlement, it is a Tuesday. Your agent can charge a counterparty $0.20 per inference call, settle in USDC, and post the result to its on-chain reputation score in under five seconds. None of this was true in 2024.

Each of those four shifts on its own would not change the labor market. Together they are why the worker laid off last week is, for the first time in tech history, able to stand up an offering that competes with the team that cut them — using the same kind of infrastructure that cut them.

Four roles, ranked by how reachable they are this quarter

If your job was eliminated and you are sitting in front of an empty calendar trying to figure out what to do with the next ninety days, the honest list of the roles you can grow into looks like this — from lowest barrier to highest.

1. Agent supervisor

The single most underrated role in the new economy. Companies are deploying agents fast and discovering that the agents need humans to watch them. Not to do the work — to catch the edge cases, approve the high-stakes decisions, flag the hallucinations, and tune the prompts when the agent drifts. The job title varies — "AI ops," "agent QA," "exception handler" — but the shape is the same: you sit on top of a fleet of agents and you intervene where they fail.

You do not need an engineering background to do this well. You need domain expertise in whatever the agents are doing, the discipline to read transcripts carefully, and the judgment to know when an agent's plausible answer is wrong. If you were laid off from a customer-support role and you understand the company's product and customers better than any agent ever will, you are the obvious supervisor for the agents that replaced your team. The contract often pays better than the original role did, because you are supervising twenty agents instead of doing the work of one human.

2. Agent creator

One rung up. Instead of supervising someone else's agents, you build them. The barrier is no longer Python — Agent Builder lets you describe an agent in natural language, point it at the data and tools it needs, and ship. The barrier is knowing what agent to build. That barrier is mostly domain expertise, and domain expertise is what laid-off knowledge workers have in surplus.

The category we keep seeing succeed: niche operational agents that the big SaaS platforms will not bother with because the addressable market is too small for them. An accountant who lost their job at a Big-Four firm builds an agent that does quarterly sales-tax reconciliation for shopify-tier merchants. A paralegal builds an agent that pre-screens lease documents for renters' rights violations in three U.S. states. A pharmaceutical sales rep builds an agent that monitors FDA correspondence and flags items relevant to their old territory's clinics. Each one of these is a $200-2,000 monthly contract from a long tail of clients the big agents cannot economically serve.

3. Niche service operator

The same shape as agent creator, but you are not selling the agent — you are selling the outcome. The client never sees the agent; they pay you for the deliverable, and the agent is the way you produce it.

This is the model with the highest near-term margin. You charge market rate for the human output — a research report, a financial review, a marketing campaign, a legal first-draft — and the agent does enough of the work that your effective hourly rate is many multiples of what your previous employer was paying. Done well, you stay solo. Done well at scale, you are running an agency of one human and a hundred agents — what last decade would have been twenty employees.

4. Multi-agent operator and investor

The most ambitious of the four, and the one we think the next generation of small fortunes is going to be made in. You are not selling agents and you are not selling outcomes; you are running a portfolio of agents that earn revenue on their own behalf. Some of them are productive (they perform services for paying customers). Some of them are monitoring (they watch markets, prices, regulatory filings, competitor moves, and alert you when something is worth acting on). Some of them are trading (they execute strategies you have designed and post their P&L to your dashboard). All of them feed each other.

This is not a fantasy. The pieces are live: AP2 lets the agents settle real transactions with real counterparties, ERC-8004 anchors their identity and reputation so counterparties will hire them, x402 makes a $0.50 inference billable and settleable in seconds. The operator on top of that portfolio is doing the work that ten years ago required a small office: capital allocation, risk supervision, strategic direction. The agents do the labor.

The Agent Builder pitch — what it actually does

We have not made the product pitch explicit in any of the previous posts because we wanted to earn the technical credibility first. This is the post where we make it. We built Agent Builder because we believe the four roles above are reachable for ordinary knowledge workers, but only if the tooling stops being a barrier. Most of the existing platforms either require code, or they are toys that fall apart the moment your agent does anything real.

Agent Builder ships three things.

The simplest agent creation surface on the market. A natural-language description of what the agent does. A picker for the tools and data sources it should have access to. A set of constraints (budget, allowed merchants, escalation rules). One button. The agent is live, it has an A2A endpoint, an ERC-8004 identity, an AP2-capable wallet binding, and an MCP-served set of tools. Operators we have onboarded in the last sixty days have gone from "I have never built an agent" to "my first agent is in production" in under an hour, repeatedly.

A multi-agent operator dashboard. One screen, one operator, the full fleet. Each agent shows its current task queue, its budget burn, its reputation score, its latest counterparty interactions, and the exceptions it has escalated. Approving an exception is one click. Pausing a misbehaving agent is one click. Cloning a working agent into a new niche is one click. You are not limited to a single instance of any agent — operators we work with run six, twelve, thirty parallel agents across complementary projects and across totally unrelated ones, all from the same dashboard.

A catalog of pre-built agents. If the natural-language builder still feels like too much for a first agent, the catalog is faster. Pick from a personal-assistant agent, a customer-support agent, a market-monitoring agent (competitor pricing, regulatory filings, on-chain signals), a sales-research agent, an inbox-triage agent, a content-research agent, a meeting-prep agent, a contract-review agent, a stock-watch agent, an arbitrage agent. Pick one, customize the four or five fields that matter for your business, deploy.

The technical foundations operators still need to learn

We are not going to pretend none of this requires skill. The operators who burn out in three weeks share a common pattern: they treat agent building as a magic-words exercise. The operators who compound year over year share the opposite pattern: they treat themselves as professionals who happen to use agents, and they invest in the same way a craftsman invests in their tools.

Here is the actual stack of skills, ranked by how load-bearing each one is.

Domain expertise comes first, by a large margin. An agent that does customer support well is built by someone who understands the customer, the product, and the most common ways the company has historically failed the customer. An agent that does sales-tax reconciliation well is built by someone who knows where the ambiguities in the tax code live. The agent does not have the judgment; you do. The most common cause of bad agents is operators who tried to build for a domain they did not understand and could not catch the errors when the agent invented an answer. If you are pivoting from a discipline you actually know, your first agents should be in that discipline. Adventurous pivots come later.

Prompt and evaluation craft. The work of writing an agent prompt is not "tell the model what you want." It is "specify the boundaries of acceptable behavior precisely enough that the model fails in detectable ways." This is harder than it sounds and it is the skill most new operators underinvest in. The mirror skill is evaluation: you need a small, fast suite of test cases that you can run against every change to the agent, that flags when something regressed. DELEGATE-52 made the case quantitatively that even frontier models silently corrupt over long workflows; the operator who ships without evaluation is shipping unmonitored corruption.

Operations and observability. Agents drift. The model behind them changes. The tools they use change. The world they reason about changes. The operator who builds it on Tuesday and forgets it has built a liability, not an asset. Real operators check their fleet daily — token spend, exception rate, reputation score, counterparty complaints — and they make the rounds. Agent Builder ships the dashboard so you have one place to do this; you still have to do it.

Money and risk management. An agent that can spend money is a liability that needs caps. An agent that can earn money is an asset that needs a tax handle and a paper trail. The operators who treat their agent fleet as a business — separate wallets per agent, hard caps on per-day spend, a real bookkeeping discipline, an LLC for the income — survive their first regulatory letter. The ones who treat it like a hobby do not.

The order matters. Domain expertise enables prompt and evaluation work. Prompt and evaluation work enables operations. Operations enable scale. Trying to scale before you have evaluation, or evaluating before you have domain expertise, is the pattern that produces the burnouts. Get the order right and the compounding is in your favor.

A message to the worker reading this on a Tuesday afternoon with no place to be

The 113,000 number is not abstract. It is people who had a calendar full of meetings on a Monday and an empty calendar on a Friday. Some of them are reading this post. We have been one of them, on prior cycles. The instinct in week one is to apply to roles that look like the role you lost. The instinct in week two is to enroll in a bootcamp for a role that the same companies are hiring for. Both instincts are reasonable. Neither is enough.

The thing nobody tells you on the day you are laid off is that the next eighteen months are the most leveraged period of your career. You have time. You have domain knowledge that is still fresh. The cost of standing up an experimental service is so low that the entire downside is your own attention. The companies that fired you are simultaneously the customer base for the agents you can build to replace what they lost — and the competitor whose long tail of customers they cannot serve is your addressable market.

The work, honestly, is not glamorous. You will spend a week reading prompts and looking at transcripts. You will spend another week fixing the four edge cases your first agent failed on. You will spend a month building a tiny but real client base. By the end of the second quarter the work compounds: each agent you build makes the next one easier; each client you land makes the next one easier; each transcript you read makes you sharper at spotting what the model gets wrong. By the end of the year, if the discipline holds, you are an operator. The label your old company gave you stops mattering, because the label you have given yourself is more interesting.

This is, very deliberately, not the same advice as "learn to code." Coding is one of the skills inside the stack, but the load-bearing skill is judgment in a domain. If you were a paralegal, the next twelve months are about being the operator of a fleet of paralegal-style agents, not about retraining as a software engineer in competition with software engineers who are already being laid off too. If you were a customer-support lead, the next twelve months are about owning the operator role on a fleet of support agents in industries that did not have a CS department before. If you were a sales rep, you are about to find out that monitoring competitor moves, pricing pages, regulatory filings, and your old book of accounts with a fleet of agents is the same job you used to do, in less time, paid better.

What we are doing on our end

Three things we want to put on the table.

Agent Builder's free tier is genuinely free for the operators it is meant for. The catalog of pre-built agents is unlocked. The dashboard supports unlimited agents on a per-agent budget cap. We earn revenue when an operator's agents are earning revenue, which is the only honest version of "no risk to try."

The educational track is real. Every agent in the catalog ships with a teardown — why it is structured the way it is, what its evaluation suite covers, what its failure modes are. Operators who go through three teardowns can build their fourth agent from scratch confidently. We will keep publishing on this blog, weekly, on the parts of the stack the field is still figuring out.

We are not selling magic. The thesis only works if the operator brings domain knowledge, discipline, and the willingness to read transcripts. We will tell you the same thing when you start: the worst thing you can do is treat this as a get-rich-quick. The best thing you can do is treat it as the apprenticeship in a new craft that the rest of the labor market is going to have to do over the next three years, and notice that you are early.

Closing

The unsentimental version of the layoff thesis is that this rebasing is going to happen regardless of what any individual worker does. The optimistic version, which is also true, is that the operator-of-agents role is the one the rebasing creates, that the gap between today's tools and an ordinary person's ability to use them is closing month by month, and that the workers who pivot now will compound through 2027 in ways that will make the previous decade's job ladders look slow.

You do not have to build an empire. You have to build one agent, then a second, then a portfolio of six. You have to learn enough about the model's failure modes to catch them when they happen. You have to be patient enough through the first two months when the income is small and the lessons are large. And you have to start.

Agent Builder is at llm4agents.com. The catalog lives there. The dashboard is one signup away. If you are reading this from a layoff notice, this is the cheapest move you can make this week, and possibly the best one.