The pitch is that you describe an outcome and the agent handles it: research the market, write the emails, run the store, make the money while you sleep.
The reality, having run one as a daily driver rather than a demo, is narrower and more interesting than either the hype or the backlash.
What actually happened
We used a general-purpose agent platform alongside a normal AI subscription for a month, on real work. The honest report:
The integration cost was the first surprise. Connecting the tools it needed took real setup time — and the monthly cost landed at roughly what a fully-functioning assistant subscription costs anyway. You are not buying capability more cheaply. You are buying the same capability and doing the plumbing yourself.
Web search broke repeatedly. Not catastrophically, just often enough that troubleshooting it became part of the routine. Which is the whole problem: an agent whose failures you have to diagnose is not saving you time, it is moving your work from doing the task to supervising a thing that does the task badly.
Memory was poor enough to erase the benefit. The theoretical advantage of an agent is that it carries context so you do not have to re-explain. When it does not, you are re-briefing constantly, and re-briefing is most of the work.
Net result: more problems, not fewer. That is the honest summary.
The thing nobody factors in: ecosystem size
This turned out to matter more than model quality, and it is worth stating because it applies to every tool choice you make.
When something breaks in a widely-used product, someone else hit it first. There is a thread, a workaround, a fix already shipped. When something breaks in a niche agent platform, you are the QA department. You find it, you diagnose it, you wait for it.
The same bug costs ten minutes in a big ecosystem and an afternoon in a small one. Over a month that difference dwarfs any feature comparison — and it is invisible when you are choosing, because nobody lists “problems get fixed before you notice them” as a feature.
So why do coding agents work?
Because a coding environment has two properties almost nothing else has.
Verification is instant and unambiguous. Code compiles or it does not. Tests pass or fail. The file has the text or it does not. The agent gets a hard, immediate signal after every single step, so a mistake surfaces in seconds rather than propagating quietly through nine more steps.
Mistakes are cheap to undo. Version control means the worst case is reverting. Nothing was sent, nobody was emailed, no money moved.
Now hold that up against where agents are usually pitched — inboxes, CRMs, browsing, outreach, purchasing. In those environments there is no test suite, no diff, no revert. The agent cannot tell whether it did the right thing, and neither can you until later, and a mistake is an email a customer has already read.
The predictor is not how clever the agent is. It is whether the environment gives fast, unambiguous feedback, and whether errors are cheap to reverse. Software development scores highest on both, which is why it is the one domain where agents genuinely landed.
“Agents that make money”
Worth addressing directly, because it is what a lot of people are actually looking for.
Making money autonomously requires exactly what agents are currently worst at: acting in the world, where feedback is slow and ambiguous, and where mistakes cost real money and cannot be reverted. An agent that misjudges a purchase, sends the wrong outreach, or prices something badly does not get a failing test. It gets a consequence.
Agents can make you money the way any good tool does — by making you faster at work you were going to do anyway. What they cannot yet do is be the business. Anything selling that is selling the demo.
Where they are genuinely worth it now
- Software development. The one unambiguous win, for the reasons above.
- Anything with a test. If you can write a check that says “this worked”, an agent can iterate against it. Data transforms, migrations, format conversions, bulk edits with a verifiable end state.
- Reversible bulk operations. Renaming, tagging, reorganising — where undo exists and the blast radius is known.
- Research with a human reading the output. The verification step is you, which is fine, as long as you actually do it.
Where to wait
- Anything that sends, publishes, buys or commits on your behalf without review.
- Anything where being wrong is expensive and you would not find out for days.
- Any workflow whose value depends on the agent remembering context across sessions, until memory is genuinely solved.
- Niche platforms, unless the capability is unavailable elsewhere. The ecosystem tax is real and it is charged in your evenings.
The honest summary
Agents are not a gimmick. They are a real capability that currently works in a much narrower band than the marketing implies, and the band is defined by the environment rather than the model.
Before adopting one, ask two questions: can it tell immediately whether it succeeded, and can I cheaply undo it if it did not? Two yeses and it will probably work. One no and you are volunteering to supervise something that was supposed to save you time.
Method
Based on a month of daily use on production work in mid-2026, alongside a conventional assistant subscription as the control. Findings are specific to the platform tested; the ecosystem-size effect is general, and the feedback-loop argument is structural rather than a comment on any one vendor.
