Intent Architects and the End of Codebases

James Turner

I was sitting in my editor late last night, watching a model refactor an entire background worker service, and, uh, it hit me that I was barely reading the code anymore. I, I, I was just reading the reasoning traces. And then this paper drops from September 2026 titled The End of Software Engineering, and it basically puts a name on that exact eerie feeling.

James Turner

The authors lay out this wild framework. In Era 1, you installed software locally. In Era 2, SaaS brought everything into the cloud. But now we are entering Era 3, which they call Agent as a Service. And the core argument is simple, but pretty provocative. In traditional software, code is the product. You write it, you ship it, you maintain it forever. In Era 3, the agent IS the software. Code gets generated on the fly, executed once to perform a task, and then completely thrown away. No permanent codebase sitting around accumulating technical debt. Just an agentic reasoning loop doing the work in real time.

James Turner

It sounds like sci fi until you actually load up the new Claude Fable 5.1 inside Cursor. Cursor benchmarked it on CursorBench 3.2, and it hit 73.4 percent, which beats out every other model they have tested. But the headline number is not even the wild part. What makes Fable 5.1 feel totally different in practice is that it runs self verification loops. Most models write a block of code, hand it to you, and just hope for the best. Fable 5.1 executes what it wrote, catches its own runtime errors mid flight, fixes them, and keeps going until the task is actually finished. Oh, and context cache reads are 75 percent cheaper than Fable 5, so you can feed it immense context without burning through your API budget.

James Turner

So I tested this on a messy multi step refactor in my own stack. I gave it the high level goal, and I watched it spin up these temporary scratchpad scripts, test edge cases, patch its own bugs, and then vanish those intermediate scripts into thin air once the job was done. It was... honestly kind of uncanny. You are watching a machine write, verify, and discard code faster than a human could even trace the execution stack.

James Turner

Which brings us to the real engineering debate here. The paper argues your role flips from line by line developer to what they call an intent architect. Less time fighting syntax, more time specifying goals and orchestrating how agents coordinate. On paper, disposable code sounds amazing, right? No legacy debt, no outdated dependencies, no refactoring backlogs. But wait. What happens when you rely entirely on uninspected, ephemeral reasoning loops? If no human is auditing the underlying mechanics because the code gets incinerated after execution, are we just setting ourselves up for silent, systemic architectural rot? That is the tension every tech team is going to have to navigate over the next year.

James Turner

Now, if you are an intent architect trying to build these custom agentic workflows, you run straight into the classic bottleneck: training data. Creating datasets is still the most manual, painful part of building specialized AI. Normally you take whatever messy data you happen to have, reshape it, label it, and train a model that ends up constrained by those exact historical flaws.

James Turner

Adaption just launched a tool called Invent a Dataset to flip that entire process on its head. Instead of starting with existing data, you start with the objective. You describe the target behavior in natural language, and the system defines the structure and generates synthetic training examples from scratch. They pair it with AutoScientist to co optimize the dataset and the training recipe together. And across their evaluations, AutoScientist outperformed human configured training setups by an average of 35 percent. That is a massive leap for teams building targeted, smaller models on custom signals.

James Turner

But then, right on cue, Alibaba comes in from the complete opposite direction with Qwen 3.8 Max 0902. This thing is an absolute monster. We are talking a 2.4 trillion parameter model with 95 billion active parameters and a 1 million token context window. They did targeted post training for coding and collaborative work, and the benchmark scores reflect it. It scored 93.0 on PaperBench, beating GPT 5.6 Sol and Claude Fable 5, and an 86.1 on OSWorld Verified. And the API price? Just $2 per million input tokens and $6 per million output tokens.

James Turner

With a 1 million token window at two dollars a million, you can practically dump an entire enterprise repository or research archive straight into prompt context and let the goliath reason over it directly. Which raises a really sharp analysis point recently published by ArtificialAnalysis. They called out how those popular intelligence versus cost plots can be super misleading because they use a logarithmic scale on the cost axis. A log scale flattens the visual difference, making multi trillion parameter API calls look right next to smaller models, when in reality the actual price gap between cheap local models and massive frontier endpoints is colossal.

James Turner

So here is the strategic crossroads for technical leaders right now. On one side, you have the synthetic data engine route: using tools like Invent a Dataset to train lightweight, highly specialized open source models that run cheap and fast on local hardware or private clouds. On the other side, you have the brute force context route: relying on goliaths like Qwen 3.8 Max, where you pay for huge 1 million token context windows to do the heavy lifting without ever touching a training pipeline. Will synthetic data engines make fine tuned small models the default, or will giant multi trillion parameter models render custom fine tuning obsolete? I think we are about to see both sides fight it out in production. That is all for today, talk soon.