# The evolution of programming languages in the AI era

> AI Release · @ai_release1 · https://ai-release.org/guides/perevod-razvitie-yazykov-programmirovaniya-v-eru-ii_en.html

Programming languages used to change through standards committees and syntax debates. In the AI era they change through coding agents, prompts and desktop tools that write and edit code for us.

## TL;DR

- The real shift is not new syntax but new workflows: AI coding agents, sessions, prompts and files now live next to the code.
- Vibe-coding turns personal experiments into work tools fast — one employee shipped an app in an evening, and within a month half the department was using it.
- Control is the weak spot: teams often do not know who updates an internal AI tool or who is responsible when it breaks.
- Model choice is still a compromise between quality, speed and money — the weekly digest names GPT-4o, Claude 3.5 Sonnet and Llama 3.1 405B.
- Agent memory and self-modifying prompts are the next frontier, and both need trust controls.

## From syntax debates to agent sessions

For decades a new programming language meant a new syntax and a new compiler. Today the visible change is elsewhere. Developers spend more time managing sessions, prompts and files, and less time writing boilerplate by hand.

That is why desktop companions appeared. In September 2026 a project called CodexDesk was published on GitHub — a desktop companion for AI coding agents and Codex workflows. It gives a local interface for managing sessions, prompts and files right on the desktop, so instead of a bare terminal the developer gets a visual workspace.

The practical side of this shift is covered in our guide [AI for programmers: coding assistants in 2026](https://ai-release.org/guides/ii-dlya-koda.html?utm_source=tg&utm_medium=channel&utm_campaign=guide_inline&utm_content=guide_to_guide).

## Vibe-coding: fast tools, missing governance

The classic story looks like this: one employee built an app in an evening, and a month later half the department used it. That is normal for vibe-coding — personal tools quickly become working tools.

The problem is that nobody controls them. Who updates the tool? Who is responsible if the data is wrong? Recent discussion of this exact case put the question of control first. A short checklist helps teams stay safe:

- [ ] Name one owner for every internal AI tool.
- [ ] Record which model and which version the tool uses.
- [ ] Decide who is allowed to change prompts and code.
- [ ] Add a human check before the agent writes to production data.
- [ ] Keep logs so you can always answer "who is responsible".

## Agents that show their work

Agents that hide their work feel broken. A Telegram bot that compares two microphone models searches for reviews, reads pages and checks feedback — while the user stares at an empty chat.

The fix is interface work: statuses for each tool, small animations and a Stop button. This turns a black box into something a normal person can trust.

If you are building this kind of automation, start with [AI agents for task automation: a guide](https://ai-release.org/guides/ii-agenty-dlya-avtomatizacii.html?utm_source=tg&utm_medium=channel&utm_campaign=guide_inline&utm_content=guide_to_guide).

## Choosing a model: still a compromise

The channel's weekly model comparison says the same thing every week: selecting a model is almost always a compromise between quality, speed and money. In that digest, three names dominate practical tasks.

| Model | Key difference | Price/access |
|---|---|---|
| GPT-4o | Called the most balanced option for practical tasks | Not stated in the digest |
| Claude 3.5 Sonnet | Named in the same top three for practical tasks | Not stated in the digest |
| Llama 3.1 405B | Named in the same top three for practical tasks | Not stated in the digest |

For a wider view of options, see [Best AI models and neural networks in 2026](https://ai-release.org/guides/luchshie-nejroseti-i-modeli-2026.html?utm_source=tg&utm_medium=channel&utm_campaign=guide_inline&utm_content=guide_to_guide).

## Memory, trust and self-modifying prompts

Agent memory is now a research topic of its own. In 2024 New York City launched the MyCity chatbot for small business owners. That case is now used to study a trust controller for LLM agent memory — and the authors show how such a controller is built and why it can be inaccurate.

Self-modification is the other edge. According to the channel, Claude can now change itself, including its own system prompt, and users can install mods to see what that gives in practice.

Meanwhile, projects like MWS Data Scout use LLMs to describe thousands of database tables. Its product owner, Sergey Tretyakov, published a Habr article explaining how that process works.

## What to watch in late 2026

The 02.10.2026 AI digest reported that Google showed Gemini 4 Argon, a new flagship model aimed at competing with OpenAI and Anthropic. The 19.09.2026 digest tracked the same race across neural-network news.

At OpenAI DevDay 2026 the company showed an updated demo built around dots. Small details like this often become the next default in developer tooling.

## FAQ

**What is vibe-coding?**
It is building working tools quickly with AI help, often by a single person. In the channel's example, one employee made an app in an evening, and half the department used it a month later.

**What is CodexDesk?**
A desktop companion for AI coding agents and Codex workflows, published on GitHub in September 2026. It offers a local interface for sessions, prompts and files instead of a bare terminal.

**Which model should I choose: GPT-4o, Claude 3.5 Sonnet or Llama 3.1 405B?**
The weekly digest frames the choice as a compromise between quality, speed and money, and calls GPT-4o the most balanced of the three. Prices are not given in that comparison.

**Can an AI agent change its own system prompt?**
According to the channel, Claude can now change itself, including its own system prompt. That is exactly why ownership, logging and trust controls matter more than ever.