Managing developers and AI agents: what changed in cost and speed
How I split work between developers and AI agents, where cost and speed really changed, and what stayed the same. Examples from systems I built in Iraq.
In short
- Writing code got faster. The time saved went into writing clear tasks, checking and review.
- I cannot give you one figure for cost savings, because the cost items themselves changed: writing is cheaper, while planning, testing and review are still expensive.
- The agent never works on real data or on the live system, and a person is always the one responsible.
In this article
I used to start every project with two questions: how many developers do I need, and how many months will it take? These days I ask something else: how many tasks can I describe clearly, and who is going to check the result?
I now work with human developers and AI agents on the same job. I am not speaking in theory: I built the Madar platform this way. Here is what changed in cost and speed, and what stayed as it was.
Where speed really went up
In writing the code itself. Madar has a backend, a storefront, a merchant dashboard, an admin dashboard, a phone app and a marketing site. I built all of it with my agents, and kept the review and the decisions for myself.
The Dijlah Engineering website and its editing panel were built the same way.
3to 4 agents
running at once, never more
Two pieces of evidence
for the cause, before we fix any fault in a live system
Outside production
is where the agent always works: an isolated copy, no real data
The speed I mean is the speed of writing. The project as a whole does not get delivered faster by the same margin, and I will explain why.
Why delivery does not speed up by the same margin
In a live system with real users, writing the screens quickly is not enough. Old versions of the app stay on people’s phones and keep talking to the same server, so every change on the server has to remain compatible with them.
That kind of work does not get faster just because the writing did. It gets faster only as far as the tests let us trust a change. An agent can write the tests too, but deciding what must be tested still falls to a person who knows where a mistake would hurt.
Where nothing changed
Understanding has not changed. I need to know the client’s problem before I ask an agent for anything, and nobody can know it for me. On a live system I inherited, I spent time on a written audit before touching anything, and it was the most important thing I did.
Responsibility has not changed either. If a mistake reaches production, the agent cannot be held to account. I can. So I put checking and review in the hands of someone independent of whoever did the work. On Madar I ran security reviews and closed what they found, because fast code needs a slower eye.
The agent writes fast, but whoever delivers the project is the one who answers when the code is wrong.
I apply the same principle when AI is part of the product itself. For Enjazo I am building a test set in Iraqi Arabic and using it to measure the quality of the replies before I promise anyone anything.
How I split the work
I never ask an agent to ‘build me a system’. I write a plan, define the contracts between the parts, then hand out small tasks. Each task comes with its own files and a check that decides whether it passed.
- I define the problem and the constraints
- I write the plan and the contracts
- An agent carries out one limited task
- Checks and tests decide
- Independent review
- Merge and delivery
Every repository has a file of binding rules that each agent reads before it starts. There is also a product file and a decision log. On Madar I wrote down the architecture decisions along with their reasons, so that a new agent does not reopen an argument that was already settled. Scripts keep the contracts and the design in sync across the platform’s repositories, so an order is defined the same way in the storefront, the dashboard and the app.
The stronger model plans and reviews. The cheaper one carries out the clear tasks, and a task moves up to the stronger model only when a check fails. Each agent works in its own working copy, so their files never collide.
One rule took me a while to learn: if the same task fails twice, even with the stronger model, the fault is in my plan and not in the agent. I split the task again rather than try a third time. And if failures are common across the board, my tasks are vague.
Working with the human team itself, in the office or remotely, is the subject of a separate article.
The new limit: what you can review
What holds me back today is how many tasks I can review properly in one day, not how many agents I can run. If I ran ten agents and went through their output with tired eyes, I would be adding more mistakes than progress.
So I run three or four agents at a time, and add more only when their tasks touch completely separate files.
Cost: where it went and where it stayed
I cannot tell you the cost fell by half. I have no clean comparison, because nobody builds the same project twice, once the old way and once the new. The cost items themselves have changed as well. Writing got cheaper. Three things now weigh more: writing the task clearly, testing, and review.
I learned from the waste too. The biggest I saw was a fresh agent reading the whole project from the beginning, every time, before it wrote a line. I solved that by writing a summary file for each project once, which spares the agent the exploring. The second was using the strongest model for everything.
What it does well
- Writing is much faster, so I can try more before I decide
- Tests and documents that usually get put off now get written
- Independent tasks can run at the same time
Where it falls short
- The agent can be confidently wrong, so every result needs checking
- Writing a task clearly is hard work, and there is no shortcut
- Review and testing are now the most expensive part of the time
Safety rules
Warning: The agent stays out of production
When a system is live and serving real people, we work on an isolated copy of the data, with notifications, messages and payments switched off. Production itself is off limits to agents.
Our rule for any fix: nothing gets fixed until we hold two different pieces of evidence for the cause.
These rules may sound strict. But an agent acts quickly and with confidence, and its mistake on real data travels further and faster than an ordinary human one.
Online store platformMadarA platform on which an Iraqi merchant opens a store and manages orders, delivery and the collection of money from one place.
What this means for the developer and for whoever hires one
What changes for the developer on the team
The developer’s job shifts from writing every line to defining the task, reading the result and judging it. That takes more experience, not less, because someone who cannot read code well cannot review it well. It is also why I do not think a beginner can skip the basics.
The skill that gains value is explaining a problem precisely, in Arabic first. Someone who can write a clear task for an agent will write an equally precise request to a colleague or a client.
My advice to every developer is to read what the agent writes the way you would read code from a colleague you do not know well: with polite suspicion, and with questions about the hard cases. The agent does not mind questions, but it only answers the ones you ask.
What to ask your provider
If you hire a team that uses AI, which tool they use is not your concern. How they verify their work is.
Questions to ask before you sign
The short version
AI made writing faster and cheaper. It did not do the same for understanding or responsibility. If you are having a system built, ask about checking before you ask about speed, because speed without checking only gets you to the mistake sooner.
Projects mentioned in this article
Questions
Do AI agents replace developers?
Not in my work. There is less of the work that needs no judgement, and more need for someone who defines the problem, reviews the result and answers for it.
Does a system cost less because AI writes the code?
Writing costs less. Understanding, planning, testing and review cost what they did before. Ask the provider how they verify the work, not how many lines were written.
Is my data safe when an AI agent works on my project?
It is, as long as there are rules: the agent works on an isolated test copy, and your real data and your live system stay out of its reach.
How do I know a provider uses AI responsibly?
Ask who reviews the code, where the tests are, where decisions get written down, and what happens if the person who wrote a particular part is no longer around.
