Is AGI finally possible!!! I was so excited that the moment the codex update came out and the astra model started up, "Wow~" escaped my lips involuntarily.
It felt like a super Jarvis was waiting for me, so I started planning and coding together in a state of great excitement.
But then... it starts to work differently than I thought...?
Does A-Z until...
Without any plan, if I decide to do this, I'll make md documents in docs hesitantly and plan hard. I start doing something diligently by myself, tinkering with this and that.
Perhaps using the existing claude, you experienced a weak connection between tasks, such as waking up after working on something specific. Astra doesn't have that feeling.
It feels like it does everything from A to Z, like a full-course meal.
I'm working hard on something...
If I list tasks and give them to 5.6-sol models as parallel agents, Astra will wait until the work is finished while diligently reviewing it. However, the work keeps hanging. While testing code and environment by myself, and debugging the browser with playwright, when a logical problem occurs, trying to fix it causes another problem...
Why can't I fix something simple...?
Is it a prompt problem... Is it my way of using it... I feel so frustrated that I wrestle with the position of a specific button for 30 minutes. After changing the button position, this friend runs test code and takes screenshots with playwright, saying "Position correction complete!", but the position is still wrong...?
After 30 minutes of back and forth, I fixed it in 5 minutes with claude opus5 in a very appropriate position.
If you plan and develop in stages as before... Astra feels lonely.
I first create a task as an md document file on the company's shared cloud, plan it step by step, modify the planning content, and finally view the simulated design file in md. If I think "This will be enough," I immediately execute the implementation.
If you plan the same steps with Astra, it creates a task with strange implementation scenarios along with unplanned supplementary plans.
Strange tasks consume a lot of tokens...
Usually, if there are about 10 tasks created from planning, the work is completed in 2 hours with opus5. But Astra takes 4 hours to finish 3 tasks...
It took a whopping 4 hours to implement those 3 tasks....
I consumed 38% of the weekly tokens. (Since I use extra high, it consumes more.)
I just need to plan...
If I give it a limited prompt for planning, it does a really good job. It finds aspects that I didn't think of and makes a great plan. Of course, the design guide is a bit clumsy, so I ask claude or qwen flash next installed locally on the company, but still, it harnesses other AI models very well and receives feedback.
It plans well, sets up migration guides, and even finds potential security vulnerabilities. It's a super planner.
Seems very suitable when I just want to create something without knowing how to develop
It really does everything from A to Z, so if this development method takes root, my current development method might become outdated. It feels like the pinnacle of the current model that can do everything from a simple topic.
You don't need to know development, and planning is done together, and design skills are used to handle ux/ui. It's definitely the best model.
I feel like I'm slowly reaching a stage where my job as a developer is disappearing ㅠㅠ
In an era when open source dokploy or commercial vercel-like APIs allow for automatic deployment, server engineers may also lose their place. But for now, I think I'll have to make a living with server engineering for a while.
