From software engineer to agent builder: what the numbers say, and what to do

This is for the engineers who are doing fine right now. Job, decent manager, endless backlog, and a quiet feeling that the ground has moved.

It has. I have been through the offshoring panic, the no-code panic, and the 2022 freeze. This one is different. Here are the numbers, and then what I think you should do.

The numbers

Demand for developers has been below the 2020 line for more than two years, while postings across the rest of the economy recovered.

The cut is landing on the youngest. A Stanford team went through ADP payroll records and found employment for developers aged 22 to 25 down close to 20 percent since late 2022. Older developers held steady. The graduates behind them are finding the door shut.

Layoffs say the same thing from the other side. Trackers counted about 122,000 tech layoffs in 2025 and 2026 passed that before the year ended. In 2024 almost no announcement mentioned AI. In 2026, about a third do. Companies stopped hiding behind "restructuring."

And the reason is in the commit logs.

Argue about definitions if you like. I have watched my own output change and I do not need a press release to confirm it.

One more number, and it is the one that matters. Indeed's economists found US software postings rose about 15 percent in the eighteen months after coding agents went mainstream, while postings overall fell. Demand did not vanish. It moved. The roles that grew are for people who direct agents and own outcomes. The roles that went are for people who turn tickets into pull requests.

What actually got automated

The middle of the job. Read the spec, find the file, write the boilerplate, wire the endpoint, write the test, fix the lint. An agent does that faster than I do and never gets bored. That was most of a junior's first three years, which is exactly why the 22-to-25 cohort is the one bleeding.

The ends did not get automated. The front is deciding what to build and for whom. The back is standing behind it when it breaks in production with a customer's money on the line. Agents do not know your customer and cannot be held responsible.

So: are you paid for the middle, or for the ends? If it is the middle, you are competing with a subscription that costs your employer a couple of hundred dollars a month.

The shift

The engineers doing well made one mental move. They stopped being people who write software and became people who make a specific job happen, with agents as the labor.

A contractor does not sell hours of hammering. They sell a finished kitchen and hire whoever swings the hammer.

Inside a company that looks like the senior who owns billing end to end and runs three agents to implement it. Outside a company it looks better: pick a job a business currently pays a human to do, build the agent that does it, charge for it.

Engineers are the right people for this and it has nothing to do with hype. You can read an API. You know "helps with ecommerce" is not a spec and "answers returns questions and issues the refund when it qualifies" is. You know how to write the check that catches a wrong answer before a customer sees it. And you have years inside some industry. That domain knowledge is the part no model has. It is the moat.

What is missing is not a skill. It is the willingness to name a customer, name a price, and put a checkout in front of the thing you built.

Why the usual plan fails

Every engineer who has this realization opens a terminal and starts building the platform. Auth. Stripe webhooks. OAuth for Google, Slack, Notion. Chat with streaming and history. A vector store. Analytics. Terms of service. Each one is a weekend if it goes well and a month if it does not. Six months later: a beautiful repo, no customers.

I have done this twice. The platform is not the product. The agent that does the job is the product. The platform is the tax you pay to sell it.

Where AgentFront fits

Skepticism appropriate here, since we built it. We built it to stop paying that tax.

You bring the goal, the knowledge, and the tools. AgentFront runs the storefront, the chat, the checkout, and the analytics.

• Each agent gets its own storefront. Click a section, describe the change. Custom domains on Launch and up.
• More than a hundred connectors, including Google Workspace, Slack, Notion, Shopify, and HubSpot, plus any remote MCP server by URL. Writes wait for a confirmation in chat unless you turn that off.
• Your own model key. Anthropic, OpenAI, Gemini, or Kimi, at the provider's rates, no markup.
• Your own Stripe account. Customers pay you. We charge a flat plan by active customers and take no cut of your revenue.
• Analytics show the questions your agent could not answer. That is your to-do list.

Explore is $9 a month. Paid plans start with a 14-day trial. The point is not that it is cheap. It is that finding out whether anyone will pay for your agent now costs a few evenings, not a quarter.

A first month

Week one. One job, in one industry you have actually worked in. Find three people who do it today and hate part of it. Talk to them. Write the job in one sentence with a "done" condition.

Week two. Build it. Two or three connectors, the playbook, the FAQ. Run twenty real scenarios yourself before anyone else touches it.

Week three. Price against the alternative, not your costs. If a part-time assistant costs $800 a month, $99 is an easy yes. Connect Stripe, publish, send the link to your three people.

Week four. Watch the unanswered questions. Fix them. Ask the first customers what they would pay double for. That is your second agent.

Do not quit your job. Do this in the evenings. Three paying customers is the bar.

The uncomfortable part

Nobody is going to hand you a title called "agent builder" with a ladder attached. Companies will keep hiring fewer people and expecting more from each one, and the ones they keep will be the ones who made the shift on their own time.

Your skills are the right skills. Reading systems, scoping problems, knowing when something is wrong. None of that got cheaper. The typing got cheaper. If typing was what you were selling, sell something else.

The something else is close. The job you already understand, done by an agent you already know how to build, sold to a customer you already know how to find. Go pick one.