Brief IA
›

Development AI: Agents, Abstraction, and Developers

Code & Dev·Tom Levy·

Development AI: Agents, Abstraction, and Developers

Development AI: Agents, Abstraction, and Developers
Key Takeaways
1ArtCraft brings together developers who use AI to clone features of well-known applications
2Craft Anything, their agent, allows for copying software or websites without writing code, either in a clean room or through disassembly
3Automation through LLM promotes more abstract tools, a trend contested as fatalistic
4The effects on mental load, the non-determinism of outputs, and the question of the level of abstraction remain open
Why it matters : The rise of AI agents in development reshuffles the cards of the profession and questions the mastery of the software process.

AI agents capable of cloning applications already exist, such as Craft Anything within the ArtCraft collection. Beyond no-code, these tools fuel a debate about the acceptable level of abstraction and how LLMs change the cognitive load on developers. Between the promise of code automation and the risks of weakened reasoning in the face of non-deterministic outputs, software enters a zone of uncertainty.

Uncertainties about Cognitive Load and Non-Determinism

The psychological effects of using LLMs on both experienced programmers and beginners remain largely unknown. The use of agents and models imposes a different mental burden than that of traditional assemblers or compilers. The non-deterministic nature of the outputs complicates reasoning about the results obtained from the same prompt, to the point where it is not always clear whether the lack of reproducibility is a flaw or expected behavior. Some argue that it is unnecessary to understand the internal workings of the generated code, claiming that the scale and speed of work are only accessible to machines. This position, appealing under deadline pressure, is likened to a form of fatalism. Questions persist about the reality of speed gains and the proportion of LLM-driven work that results in delivered features, bug fixes, or added value—concerns that were already present before the rise of LLMs.

What Tools Are Doing Today: From Functional Cloning to Advanced No-Code

A collection called ArtCraft, designed by developers, uses AI tools to clone features of established applications, including Adobe Creative Suite, Microsoft Office, and AutoDesk CAD software. Among its projects, Craft Anything stands out as an agent capable of studying software or websites and then prompting an LLM to produce a copy. Two approaches are offered: reproducing external behavior in a clean room or disassembling binaries. The use of Craft Anything can be done without writing code, making it an extreme version of no-code.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Towards More Abstraction? Arguments and Counterarguments

The idea that LLM-driven development allows for cloning software without programming is widely accepted. This approach would encourage the creation of more abstract tools and shape generations of developers whose first, if not only, contact with programming would be through these intermediaries. One argument posits that high-level languages would be the next to be automated by LLMs: models would write the code while humans focus on architecture and higher-level ideas. Some suggest that the natural evolution of software development is towards greater abstraction and that current concerns about code generated by LLMs will seem trivial in the future. In contrast, critics label this perspective as fatalistic, advocate for the choice of abstraction level, and emphasize that adding layers makes it more difficult to identify errors and areas for improvement.

Where We Come From: From Punch Cards to Optimizing Compilers

In the early days of computing, programs were close to the hardware and used media such as tapes or punch cards. The emergence of high-level languages allowed for writing with a syntax more readable than machine code. The first compilers elicited mixed reactions to the abstractions introduced, but decades of progress have led to compilers generating machine code considered superior to manual writing. In this context, the question of the level of abstraction truly necessary for software development remains. It is suggested that the increased delegation of work to AI agents will make it more difficult to answer this question, and as practices shift towards agents, it will become harder to formulate these inquiries.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.