Clarifying a few things before any possible misunderstandings, your team will still need humans to:
- Design and create software from the first principles
- Write-review-describe code
- Solve problems
Offloading human thinking process to non-thinking machines will not work flawlessly, nothing to debate there.
Now, I can safely talk about why you might end up so slow and perhaps even unemployable as a software engineer, and how to avoid this.
models & harness
With each new release, they become better at:
- tool usage
- efficient context building
- parallelising work units
and these improvements are only accelerating.
This acceleration only widens the gap between people who can use these tools well vs people who can't.
humans & code
We as humans can't read/write two functions at the same time, yet given the correct instructions, agents can read/write tens or hundreds of them.
And this is the part where you, as a software engineer, have to adapt. You must be able to utilise them, ideally better than others.
Almost any experienced software engineer will tell you that models usually do not produce the code the way a good engineer would write, and yes, these experienced engineers are often right.
But remember, everytime we reject a certain approach, approve another way of doing things, we are helping these models to become better. You can observe this by comparing today's models to their debut or even just a year ago.
And these models do a better job when the codebase is well-structured and documented because they can follow existing conventions, thus avoiding split-brain approaches which is commonly seen weakly designed / vibed projects
Now, how do you keep up?
Being honest, unless you pay up for the frontier model's when they are introduced, you won't know what you can or can not do with those models. Thus someone who does that will have better chances of leaping forward.
Yes, all-rounder models are becoming available and cheaper, but frontier will always be the frontier AND when you are competing, you will want to have the frontier.
Leaving model quality out of the way, what can you do for parallelisation as the human?
Go back to first principles like how you would solve anything complex pre-ai. Break down into focused executable pieces as much as you can.
Instead of throwing the problem at your model and walking away, do these:
- Identify the units of work which need to be done, what does it require? At this step you must use the general purpose model to gather information only.
- Make the "necessary" decisions yourself. The convenience you give up here will pay back later.
- Ask the smartest model to see if it can raise justified objections against your decisions and go through them.
- At this point a coherent, unambiguous plan should be constructable.
- Use the smartest model/s in your arsenal to lead the smaller and efficient models for doing the work.
- Always loop your agents on the implement then review approach. Do your own review on the checkpoints where they are most important to you, not every little type or function requires a human in the loop.
- Utilise agent multiplexers like "Herdr". For instance, Claude Fable is extremely efficient for managing smaller models.
This is not a magic recipe that solves everything, you still have to be in the loop.
However, this will allow you to work on multiple things at the same time without degrading quality because you are still the main driver.
And this is precisely what most employers will want from you.
Someone they can still rely on while producing more work.
Good luck.