Table of contents
Written by Martin Ågren - August 9, 2026
Development has changed:
Software development has changed because of GenAI.
What developers say:
I have spoken with developers, engineers, and IT managers in Swedish organizations, and they describe the same trend: Development has become faster, and methods have changed.
Faster development:
One IT manager told me that software solutions that previously took several weeks to deliver can now sometimes be completed in less than a week.
AI is capable:
AI can generate, review, and improve code.
Humans are needed:
Software development expertise is still important! And human judgment and decision making is still needed.
The developer's responsibilities:
The developer still needs to guide the development process, as an orchestrator prompting, providing specs, defining goals, connecting tools, reviewing, testing and validating output, making technical decisions, and taking responsibility for the finished result.
Developers I have spoken with have described using different AI coding systems, and the reasons for choosing them:
Claude Code:
Reason: This system has been a leading system in the AI coding space for a while.
Codex:
Reason: The organization already had an agreement with OpenAI.
GitHub Copilot, in VS Code:
Reason: The organization already had an agreement with Microsoft.
Kiro:
Reason: The organization already had an agreement with Amazon.
Local inference on computer:
Reason: increased data security.
Hybrid approach. Local inference / Frontier models:
Reason: Hoping to get the best of both worlds by combining local inference with frontier models running on cloud servers, balancing capability, cost, GDPR, and data security.
What it means:
Break the project into small, clearly defined tasks. Review each result before moving to the next step.
OpenAI recommends "narrow silos":
In August 2026, I listened to an OpenAI webinar where they recommended this method which they called "narrow silos" (prompt decomposition, instead of using one large super-prompt).
Small enough task - easy to review:
As OpenAI suggested in their webinar in August 2026:
The developer should be able to read and understand what the LLM outputs. If the output is too large to review, the task was probably too broad to begin with.
Agentic workflow still possible:
The AI system can still work through several steps with high autonomy, to reach the goal defined by the developer, but within a narrow task.
Advantage:
More control:
The developer keeps control throughout the project.
Find problems earlier:
Errors and weak decisions can be corrected before they affect other parts of the project.
Easier final review:
The finished result can be easier to review because the developer has already reviewed each step.
Disadvantage:
Slower process & more manual steps:
Reviewing the work step by step usually takes longer (at least in the initial phase) than letting the AI system complete one large task from a super-prompt.
What it means:
A long-horizon task is a large task that the AI system completes over many steps. The developer gives one detailed super-prompt (or a set of instruction files, system instructions, etc) describing the final goal.
Senior developers reviewing final code:
I have heard from several developers that Method 2 results in senior developers having to review a large amount of AI-generated code that junior developers prompted the AI to generate earlier in the process.
Real-life example:
A software engineer at a large Swedish company told me that his team had tested both method 1 and 2. They were especially interested in Method 2 because it requires fewer manual steps. It was a work in progress, and they were not fully satisfied with the result, since they found that the end result needed to be thoroughly inspected and corrected by a senior developer.
Advantage:
Fewer and simpler steps initially:
The AI system can complete more of the project without constant input from the developer.
Disadvantages:
Problems can spread:
A weak decision made early can affect many later parts of the project before a developer notices it.
Harder final review:
The final code review becomes larger and more demanding.
"No fun":
Some developers dislike mainly reviewing code generated from prompts written earlier by junior developers.
They would rather take part in the creative process of building the software.
What it means:
"Cognitive debt" means knowledge or skills you have lost because you have relied too much on AI to perform work for you.
Solution - Coding without AI:
Developers may need to sometimes code without the help of AI to maintain their knowledge and skills that may otherwise fade over time.
Advantages:
No subscription + No API costs:
Local models avoid subscription fees and external API charges.
GDPR + Data security:
Source code, sensitive data, and company secrets can remain on the organization’s own computer or server.
Availability:
The model runs locally, so you are not dependent on slow or unavailable external servers.
Disadvantages:
Capability:
Local models are generally less capable than the best frontier models running on cloud servers.
Hardware:
Very capable open weight models require powerful hardware.
What it is:
A hybrid approach combines local inference with frontier models running on external cloud servers (cloud inference).
- For simpler tasks: Use local models
- For difficult tasks: Use frontier cloud models
Lower costs:
Using local models for simpler tasks can reduce subscription and API costs.
Maintain overall quality:
Frontier cloud models can be used for difficult tasks, helping maintain overall quality in the finished result.
Some AI coding systems let you choose between using a CLI in the terminal or a desktop app.
Extension:
Some developers use AI coding systems as extensions in code editors like VS Code.
CLI:
Some developers use AI coding systems through a command-line interface, often in the terminal inside VS Code.
Desktop app:
In August 2026, I listened to an OpenAI webinar where OpenAI described its hope that developers would increasingly use Codex mode inside the ChatGPT desktop app. Time will tell...