Important notice, if you want to code with AI at one of the levels, you must master all lower levels, there is no short cut. If you want Level 4 multi agent coding, you must go through Level 0, Level 1, Level2 and Level 3.
Second notice, You only reduce the coding time but software development is a little bit more, maybe around 60-80% more.
Last remark, all values in the table are good guesses based on my experiences.
Overall: now in 2026 software development works pretty well under some conditions:
High quality code base
High experienced software developers
This is the perfect combination for successful AI assisted software development. AI is getting faster and tools getting better, it's a lot of fun to code with AI. At the moment AI is in the storming phase.
You want an example, here it is a small Golf trainings app with mostly complete AI generated frontend and AI maintained and extended backend.
Bullshit Asymmetry in Software Development means that is very easy to develop a new system on a green field. You can easily go on putting new features into the code but over time, often faster then slower, you have to put more and more effort to add new features until you reached the culminating point. At the software development culminating point your resources (developers) are exhausted, no new feature can be implemented in acceptable time. Next the development team is wearing out and then I'm appearing.
Software reconstructing is more expensive then writing the software. The costs are asymmetric. Build a new software is much cheaper then reconstruction/rescuing legacy software.
But software reconstructing lowers the risk of a complete miss. The improvements are coming at a constant speed. No new teams are needed. Reconstructing software can be realized by the origin team, maybe extended by some external experts.
There is a third way, re-implementation of an existing system. This way combines low risks, like software reconstruction with the low costs of building a new software.
The efficiency in many software companies are low. You may know it, endless meetings with to much people never needed. Presence counts more then results. But here some simple things that increasing the efficiency are really easy to implement:
One meeting free day a week (Meeting-frei Freitag)
No back to back meetings, at least 5 min break between
Time boxed meetings
Regular breaks in long meetings
This ideas will not fix every thing but they are simple and easy to implement. And these ideas could be a turning point to more focused, result focused work.
Fun fact, after introducing a meeting free Friday, after some time, a lot of meetings appears on the meeting free Friday, because there was a lot of free time. But now the Thursday is now meeting free.
For all ‘Happy Path’ developers, proper error handling is a
fundamental feature of software, and it is quite possible that error
handling is for half of your code.
Expect the unexpected: not all English postboxes are red. This Victorian one stands in Dublin.
There two types of
errors:
Retryable Error
Permanate Error
Retryable error: a
temporary issue has occurred. a retry is likely successful. Example
causes: firewall reboot, db maintenance, HTTP 5XX.
Permanent error:
error where a retry will always end-up with an error. Example:
HTTP4XX, NPEs (Null Pointer Exceptions).
After a retry
sequence, Retryable errors should become permanent errors.
And at last, don’t
forget Logging and Monitoring of errors.
No. The rising of LLM will increase the demand of software developers.
Longer Answer
A LLM can generate text based on the training data. That's brilliant if you doing you homework or writing a message a teacher (South Park episode "Deep Learning"). If you search for information, LLM are brilliant, comparing things, aggregate information that the jobs LLM can help you a lot.
But Coding, writing software; is not writing text. If you want to learn a programming language, a LLM is great. If you searching for boiler plate code or a strange configuration property, LLM is my tool of choice. That's not coding. That's a advancment over Google or Stack Overflow. A LLM can improve autocomplete in my IDE (Kai Lentit on Y). LLM can't write software that is correct, that is fast, that is reliable, that is save, that is robust, that you and your business can rely on. So far, LLM has never been able to handle any programming task for me that went beyond the beginner level. And believe me, I tried it hard and I found out this a general limitation of LLM. I can explain it, but it's an other story.
LLM is great of doing Bullshit jobs. It will free people for better jobs. LLM increase the efficiency of software development in general.
LLM is a good new tool, but a fool with a tool is still a fool, with brand new legacy code. LLM is a just a tool like Excel or the Internet, LLM is just a high sophisticated tool (2026).
Summary
LLM increase the software development efficiency
LLM lower the entry level for software development
IMHO LLM will create a Rebound Effect on software development
LLM will help to create more Software faster
Software Developper will be more important than before.
Null
is biggest flaw of Java, saying some developers. Null causes NPEs
causes crashing software. IMHO this is the same vibe as C vs Rust and
I totally disagree. Crashing software is caused by developers and not
by the programming language.
The easiest and best way to prevent NPEs is … proper exception
handling. It is as simple as it sounds.
Bye the way, NPEs are often a side effect of OOP. All Objects are
unsafe until you checked them. But checks are unsafe so go with
exception handling.
For most software companies resilience means that the replication factor is 2 or higher.
Well that's far from good. You resilience is still bad as with two instances of your weak micro service.
One way to increase the internal resilience of your service is to decouple external service calls. This includes Data Base operations, Kafka, HTTP. For HTTP calls it is very common and most developer understand that this HTTP calls can go wrong but Data Base calls, yes all external services are running over a network, firewall, load balancer, switches, ... So it's a good idea to make all external call more resilient:
Put the external calls into own treads. This keeps your application running in case of an error or time out.
And have a time out or watchdog on it. By the way, Java futures are a easy way to do it.
Check the results also for write operations.
Extend logs and monitoring to recognize external call errors.
