2025 Results: AI in Development

The Future is Already Here

It seems to me that with the release of Claude 4, we passed the point where development with AI became more effective than without it. And with GPT-5.2 and Gemini 3 Pro/Flash, this became obvious.

Vibe Coding Emerged

Creation of applications by an agent, without analysis and “appropriation” of code by a human. It works quite successfully for projects up to several thousand lines. Good for concepts.

Agents Learned to Write Average Code, But Haven’t Learned to Build Architectures Yet

There are very few benchmarks evaluating code quality, and almost none for architecture.

You Have to Pay

Companies or developers themselves will have to pay for premium plans in AI-IDEs — they give more stable results. 200–300 USD/month is the new reality. Attempts to squeeze the maximum out of cheap plans are economically impractical.

Clear Task at the Input

Agents are critical to the quality of the task description. Development through a plan, specifications/requirements partially solve the problem and gradually penetrate the IDE. However, the ability to formulate a task in the required form is one of the most important factors in the quality of agent performance.

Rigid Verification at the Output

Outgoing control quality is critically important for stability. All types of checks should be included: static analysis, all levels of tests, test coverage analysis, llm-as-a-judge, code review, etc.

Forced Towards Good Architecture

Limitations on context size and model attention force a shift to modularity and loose coupling: it’s easier for agents to manage small modules than monoliths.

We Are All Architects Now

The developer still has to be an architect and operator of the SDLC process, but less and less code needs to be written.

Boom of AI-first Development Environments

So far, no terminology has been established, and everyone just talks about agents. But environments can be divided by the level of user proximity to the code:

• AI-assistant: Cursor, Windsurf, JetBrains products, Zed, IDE extensions (Claude, Codex, Warp, Trae)…
• Vibe coding (much less attention to code and more to communication with the agent): Devin, Kiro, Qoder, Claude Code CLI, Codex CLI, OpenHands (open source), agent mode in Cursor…

AI-Assistant IDEs are Evolving

In a year, several new ones appeared, and old ones advanced significantly.

• Cursor — still the leader in features (indexing for fuzzy search, different types of rules, easy rollback of edits, good instructions, “free” Grok Code type suggestions, and a built-in browser).
• Antigravity — a confident second place thanks to strong planning and the most convenient work with the plan.
• Claude Code — already a “classic” for work. The problem, for me, is the weakness of integration into the IDE compared to Antigravity and Cursor.
• Codex — a clone of Claude Code, with few features so far. Same problems.
• JetBrains products — I lack control (rollback, rules, commands), so I haven’t looked at them for a while.

Dream IDE

In an ideal IDE, I would like to see:

• local index for fuzzy search and selected documentation
• programmable workflow based on a group of agents, with separate instructions and configurable human-in-the-loop at the project level
• a visibility setting for the agent separate from .gitignore
• persistent project memory
• development through specifications/requirements

LLMs Increase Efficiency at All Stages of SDLC

LLMs one way or another increase efficiency at all stages of SDLC. From writing Product Requirements Document and Software Requirements Specification to analyzing logs and metrics in production. They don’t replace a highly skilled specialist but allow doing what previously there weren’t enough resources for.

LLMs and Agents in Infrastructure

So far, results are more modest than in code development. They handle generation of basic configurations and project skeletons well but often hallucinate non-existent resources and parameters and are prone to over-complicating solutions.

LLM is a New Tool for “Talking” to Code

Suitable for exploring legacy and complex code, onboarding new developers, searching for bug causes, and understanding non-obvious dependencies.

Product Security Crisis

Relaxation and lack of review lead to a large amount of potentially vulnerable code appearing in repositories, muted checks, and so on.

Start of Layoffs

Layoffs went two ways: replacing employees of a similar level or delegating their tasks to more qualified personnel. While at the beginning of the year companies limited themselves to hiring freezes, by the end they shifted to direct position cuts.