How to thrive in the agentic AI era
By Flavio
Tagged: AI
Agents made producing things cheap. What's scarce now is judgment, trust and knowing what done looks like. What that means for developers, professionals, prosumers and small business owners.
This site’s repository got 221 commits last week. Some were mine, most were agents I briefed, a few were agents running on a schedule while I did something else. In the last six months it got 3,588.
In the same period I built and released Mac apps for people who will never open Terminal: a flashcard printer for teachers whose backs line up with the fronts, an app that saves the voicemails sitting in an iPhone backup, another that exports every voice memo with its date and transcript. A year ago each of them would have been a month of evenings. Now the first working version takes a day, and the rest of the week goes into the parts nobody sees: the screen that tells you what to do next, the signed and notarized build, the update check, the sales page.
So I’ve had a good look at what changes when an agent does the producing. My short answer is that this is the best time in decades to be someone who cares about doing good work. The tools are the same for everyone, and they’re getting cheaper every month. What isn’t the same is the person giving the brief and checking the result.
This post is my long answer. It’s for four groups of people, because I think the advice is different for each: developers, professionals who aren’t developers, prosumers, and people who run a small business. Read the general part, then jump to yours.
Where the scarcity went
Every time something gets cheap, something next to it gets expensive. When printing got cheap, attention got expensive. When storage got cheap, finding the right file got expensive.
Agents made a long list of things cheap at once.
| Got cheap | Got expensive |
|---|---|
| A first version of an app, a page, a script | Deciding which one is worth building |
| Boilerplate code, forms, settings screens | Reading a diff and knowing it’s wrong |
| A draft of an email, a report, a contract | Knowing which clause matters for this client |
| Research summaries | Knowing which source to trust |
| Translations, rewrites, variations | Taste, and the nerve to say no |
| Running a task every night | Someone reading what the task produced |
| Building a competitor to your product | People who already trust you |
The left column is what most people talk about. The right column is where the money and the careers will be for the next few years.
Look at the right column for a minute. Nothing there is new. Judgment, taste, trust, knowing your domain, owning the outcome. Those were always the valuable part of any job. For decades we could hide them behind effort. A developer could feel productive after a day of typing. A consultant could justify a fee with a 40-page report. Now the typing and the report cost almost nothing, so what’s left is the part that was always the job.
I wrote about this from the pricing side in software was never worth the hours. Customers paid for a result they could trust, and the effort behind it was always invisible to them.
The loop that works
Every good use of an agent I’ve seen, mine and other people’s, follows the same loop.
The agent does one box. You do the other four.
The people who struggle with agents usually skip the first box or the fourth. They ask for “a landing page” without knowing what the page has to achieve, then they read the agent’s summary instead of the page. They get something that looks finished, and nobody involved knows if it’s right.
The people who thrive spend more time in the first and fourth boxes than they used to, and less in the third. It feels slower, but they redo much less work.
The last box is the one that compounds. When I correct the same mistake twice, it goes into AGENTS.md or a skill, so the next session starts from it. This repository has a CLAUDE.md with rules like “commit only your own files, because sessions share one git index” and “never publish a real IP address in a screenshot”. Each of those lines is a mistake that happened once and won’t happen again. A chat forgets everything when you close it. A file in the repository remembers.
That loop is the whole method. The rest of this post is what it looks like in four different lives.
If you’re a developer
Your output stopped meaning anything
For most of my career, how much I shipped was a decent proxy for how good I was at the job. That’s over. An agent can produce a thousand lines before lunch, and so can everyone else’s agent.
3,588 commits in six months sounds impressive, and it tells you nothing. Some of those commits are posts nobody reads. Some are features I removed a week later. The number that matters is how many of those things people use, pay for, or come back to, and that number grows much slower.
So I stopped measuring days by what got produced. I measure them by decisions. Did I decide what this app is for? Did I cut the feature that didn’t belong? Did I find the bug before a customer did?
Building is a very convincing form of procrastination now. You can feel productive for a month and create nothing of value. I wrote about this in when more people build software. A working first version used to prove something. Now it proves you have a subscription.
Review is the job
Faros AI measured telemetry from more than 10,000 developers in their productivity report. Teams with high AI adoption merged 98% more pull requests, and the time spent reviewing them went up 91%.
That’s the developer job in 2026. Production doubled and the bottleneck moved to review.
Reading code fast and knowing when it’s wrong is now the core skill. Agents write code that compiles, passes the tests they wrote themselves, and still does the wrong thing. A function catches an error and returns an empty list, so the page shows “no results” instead of failing. A webhook handler works perfectly the first time and charges the customer twice when the provider retries.
You catch those by understanding the system, not by reading the agent’s report. My rule is to read the files, not the summary. Open the page, run the build, read the diff. A pull request merged without being read is a slop grenade, and someone else will be standing next to it when it goes off.
Ask before you let it write
The best way I know to get better at review is to use the agent to understand code before using it to write code.
Point it at a codebase and ask sharp questions. What happens if this webhook arrives twice? Where does this email go after we store it? If this returns 200, did the user actually get access? Make it answer with file and function references, and separate what it confirmed from what it assumed. I wrote the method in use AI to understand code.
This works especially well if you’re early in your career. The fear I hear most from junior developers is that they’ll never learn, because the agent writes everything. You learn the same way we always did, by building a model of the system in your head. The agent can be the most patient senior developer you’ve ever had, if you ask it questions instead of asking it for code.
Move down the stack
Agents are great at the code inside a function. They’re much less reliable at everything around it: DNS, deploys, databases, auth, payments, email deliverability, backups. That’s where software breaks in production, and that’s where people who understand systems are getting more valuable.
Earlier this year I had to move my newsletter server to a new, smaller DigitalOcean droplet. Apache, PHP, MySQL, certificates, cron jobs. Codex did it over SSH in about half an hour, and it even disabled the cron job on the new box so two servers wouldn’t send the same emails. Then the migration left a couple of subtle problems that only showed up when I sent the next newsletter. Tracking them down took knowing how that server worked, which the agent’s summary couldn’t give me.
If you want to bet on one area, bet on the boring infrastructure. It’s where agents need supervision the most, and where mistakes cost real money.
Put the rules in code
When an action costs money, deletes data or emails thousands of people, “always ask me first” in a prompt isn’t enough. The model may follow it or it may not.
When I let an agent buy a domain through the Cloudflare API, the API call was a few lines. The real work was a server that enforces a $20 cap, quotes that expire in five minutes, and a confirmation form drawn by the client, not by the model. When I built a hosting cost estimator, the LLM never picks a price. A deterministic engine does the math in integer cents, and the model only writes the explanation.
Knowing where to put that boundary is a skill, and it’s a developer skill. Most teams adopting agents will need someone who has it.
Ship something of your own
This is the part I’m most excited about. The cost of building a small, useful product dropped so much that a developer with a weekend can do what used to need a small team and a seed round.
I’m not saying everyone should quit their job. I’m saying a side project is now a cheap experiment, and every experiment teaches you the right column of that table: picking a problem, pricing, writing a sales page, answering a support email. Those skills make you better at your day job too, because they’re exactly what agents can’t do for you.
If you’re a professional
By professional I mean anyone who does knowledge work and isn’t a developer. Accountants, lawyers, marketers, designers, teachers, consultants, project managers.
Your domain knowledge is the spec
An agent can draft a contract in a minute. It doesn’t know that this client got burned by an auto-renewal clause two years ago, or that this jurisdiction treats a certain term differently, or that the partner reviewing it hates passive voice.
You know. That knowledge used to be one input among many, next to the hours of drafting and formatting. Now it’s almost the only input that matters, because the drafting is free.
So the professionals who thrive are the ones who can write it down. What does a good version look like? What are the three mistakes juniors always make? What do you check before you sign? If you can write that in a page, you can hand it to an agent with every brief, and the output starts from your standard instead of the internet’s average.
In this repository that page is CLAUDE.md. For you it might be a document called “how we write proposals” that you paste at the top of every conversation.
Do it by hand three times first
The mistake I see most often is automating a task before understanding it.
Someone hears that agents can process invoices, sets up an automation, and three weeks later finds out it categorized half of them wrong. Nobody noticed, because nobody was checking.
Run the task with the agent in a chat, by hand, three times. Every time you correct it, write the correction down. When the corrections stop, you’ve got a brief good enough to repeat. Only then think about putting it on a schedule. In what are your agents doing all night? I went through what people actually run unattended, and the jobs that work all started this way.
Be the person who knows how
In every team I’ve talked to, there’s one person who figured out how to use agents well, and everyone else asks them. That person’s value inside the company went up a lot this year.
You don’t need to be technical to be that person. You need to be curious, write down what works, and share it. Turn your best brief into a template. Show a colleague how you checked a result. Write the internal page that explains which tasks you trust the agent with and which you don’t.
Be careful with what you paste
Client data, contracts, medical records and salaries don’t belong in a random chat window. Check what your company allows, check what the provider does with your data, and if the answer is unclear, don’t paste it.
For sensitive work there’s a real alternative now. Models you can run locally on a recent laptop are good enough for summarizing, extracting and drafting, and nothing leaves the machine.
If you’re a prosumer
By prosumer I mean the person who’s not a developer but uses their computer seriously. You have opinions about note-taking apps, you’ve tried three task managers, you automate things with Shortcuts, you edit your own videos.
This might be the group that gained the most.
You can build your own tools now
For years, if your tool didn’t do the thing you needed, you had two choices. Find another tool, or live with it. Now there’s a third: describe what you need to an agent and get a small script or app that does exactly that.
A tool that renames your photos by the date they were taken, a page that turns your reading notes into flashcards, or a script that checks every morning if a sold-out product is back is an afternoon with an agent, and the result fits you exactly.
I’ll be honest about something here. Many of the Mac apps I sell are things a determined prosumer could build with an agent over a weekend. I sell them anyway, and my bet is that most people want the result tonight, signed, tested and with someone to email when it breaks, more than they want a weekend project. Both choices are good. What’s new is that you get to choose.
Keep your stuff in files you own
Agents work best with data they can read. Plain text, Markdown, CSV, a folder of documents, an app with a command line tool.
I manage my to-dos in Things, and agents add reminders there through its CLI. Every post and every operational doc of this site is a Markdown file in a Git repository, so any agent can read years of decisions in seconds. When your life is locked inside an app with no export and no API, an agent can’t help you with it, and every new tool you try starts from zero.
So when you pick a tool now, I’d add one question to the list: can an agent read and write my data in it?
Watch the subscriptions
There’s a new kind of pressure. A $200 a month plan feels wasted if you don’t use it, so you find something for it to do. I saw this all over the replies to the “agents running all night” question: people running agents to justify the plan, producing reports nobody reads.
I pay $100 a month for ChatGPT Pro, and when I priced my usage at API rates it covered about $745 of work in 30 days, all of it daytime work I needed anyway. Start with a $20 plan, and move up when you hit the limit on real work, not on work you invented to use the limit.
If you run a small business
I run a small business. I sell apps, I publish this site, I send a newsletter, and I keep a set of free handbooks up to date. So this section is the one closest to my own daily life.
Everything you sell got easier to copy
You can now do things that used to need hires. A sales page, a support macro, a translation, an internal tool, a small app. I do all of these with agents, and I would have needed at least three people for them five years ago.
Your competitors can too, and so can anyone who wants to become your competitor. If what you sell is mostly the effort of producing it, the price is going down. A logo, a basic website, a standard report, a generic course.
I saw this with my own courses. I had six free AI courses on this site, with more than a hundred lessons between them. A single post on the same subject got far more readers than all of them together. When anyone can generate a course in an afternoon, a course stops being special. So in November I retired all six and merged them into one excellent post per subject. Less content, better content, easier to find.
Trust is the moat
What’s hard to copy is everything in the bottom row of the table: people who already trust you.
Anyone can generate a newsletter in a minute. Nobody can generate the years of people opening mine because the last issue was useful.
For a small business this means investing in whatever makes people pick you without comparing: a reputation in your town, a real person who answers the phone, a guarantee you keep, a brand that stands for something. I wrote about this in software with a point of view: when everyone can build the generic version, the opinionated one wins.
Agents are becoming your customers
This one is coming fast. People are starting to ask an agent to find a plumber, compare three hosting providers, or buy a gift. The agent reads your website, not your Instagram.
If your prices, opening hours and services are in a PDF, or in an image, or behind a form, the agent might skip you. Plain text pages with clear prices and clear policies are suddenly a competitive advantage. This site serves an llms.txt file for the same reason. I covered the shopping side in what happens when AI agents shop for us.
Automate the boring work, keep the relationships
The best uses I’ve found in my business are boring. An agent checks the provider prices on two of my sites every two weeks and opens a pull request I read and merge by hand. Another adds internal links between posts every morning. Another records the day in the changelog every evening. None of them talk to a customer.
When a customer emails me, I answer. An agent can help me find the order or draft the steps, but the reply is mine. A small business wins against a big one because a person cares, and handing that to an agent throws away the one advantage a big company can’t buy.
Count what it costs
Tokens are a cost of goods now. Agents running on schedules, tools with AI features, subscriptions for three different assistants. It adds up quietly.
Once a month I’d look at every scheduled agent and ask one question: did its output change a decision this month? If not, turn it off. A report that you skim and ignore is a cost with no return.
What all four have in common
Looking back at the four sections, the advice keeps landing on the same few habits.
| Habit | Developer | Professional | Prosumer | Small business |
|---|---|---|---|---|
| Know what done looks like | Tests and a clear spec | Your standard, written down | What the tool must do for you | What a customer must get |
| Check the result yourself | Read the diff, run it | Read the draft as the client would | Try it on real data | Read the page as a customer |
| Write corrections down | AGENTS.md, skills | A “how we do it” document | A note with your best prompts | Playbooks for each task |
| Keep a human where it matters | Payments, auth, deletes | Signatures, advice | Your money and accounts | Every customer conversation |
| Own something agents can’t copy | Systems knowledge | Domain judgment | Your data, in your files | Trust and relationships |
None of these is a prompt trick. Prompt tricks expire with every new model. These habits got more valuable with every new model I tried.
The worries are real
I’m very optimistic about where this goes, and I still don’t want this to read like an ad for agents. A few things worry me, and each one comes with the reason I’m hopeful anyway.
The juniors. Many companies hire fewer of them because agents do the junior work, and every senior developer was a junior once. That’s a real problem for the industry to solve. But a curious beginner today has a patient tutor available at 3am and can ship in a month what took me a year when I started. The path into the job goes through building real things in public much more than it used to, and the people who take it will get good very fast.
The joy. A lot of developers loved the typing, the flow of an afternoon writing a feature, and agents took exactly that part first. I found my joy again one level up, in seeing ideas become real things in a week, and I wrote about it in AI and the joy of programming. I know that’s not true for everyone. What I see every week, though, is people who never wrote code feeling the joy of making a tool work for the first time, and there are millions more of them than there are of us.
The noise. When producing is free, there’s more of everything: apps, posts, newsletters, courses. Getting attention gets harder for everyone, me included. The flip side is that careful work stands out more against a background of generated sameness, and a voice people trust becomes worth more every month.
And the prices. I don’t know what small software is worth when anyone can generate a first version, and I’m watching my own sales like everyone else. But cheaper software also means problems that were too small to solve are now worth solving. A flashcard printer for teachers would never have justified a month of evenings. In a few days of work it’s a real product, and the teachers who need it finally have one.
What I’m betting on
Advice is cheap, so here’s what I’m doing with my own work.
I’m building for people who will never use an agent. My newest apps are for teachers who print flashcards and for people with years of voice memos and voicemails stuck on their phone. They won’t build these tools themselves, and there are far more of them than there are developers.
I’m publishing less of what an agent could write for anyone, and more of what only I can write, built on my own runs, numbers and mistakes. That’s why the six courses became posts, and why this post quotes my own commit history instead of a survey.
I’m writing every correction down. The CLAUDE.md and the skills in this repository get better every week I use agents, and they’re the one thing I’ve built this year that I’d never want to start over.
And I’m keeping my own hands on anything that costs money or that a customer reads from me. Agents prepare a lot of it. I approve all of it.
Some of these bets will be wrong, and I’ll write a follow-up in six months saying which.
But I’ve been building software for more than twenty years, and I’ve never been this excited about what one person can do. The tools got better than anyone predicted, and they’ll keep getting better. The hard parts that are left are judgment, taste and caring about the people you build for, and those were always the best parts of the job.
Want me to talk about your product? You can sponsor this site.
Related posts about ai: