DEV Community

Cover image for How AI IDEs Actually Work - Under the Hood
Prajwal Gaonkar
Prajwal Gaonkar

Posted on

How AI IDEs Actually Work - Under the Hood

#architecture #llm #tooling #ai

When we ask an agentic IDE like antigravity to “explain this” or “write code like this”, what actually changes?

And how does it return exactly what we asked for?

Let’s break down what’s happening under the hood.

Overall Workflow 

User Prompt
↓
Context Builder (files, code, selection, search)
↓
LLM (predicts next action)
↓
Tool Call (if needed)
↓
Execution Layer (file update / command run)
↓
Result returned
↓
LLM again (decides next step)
↓
Final response / more actions
Enter fullscreen mode Exit fullscreen mode

1. It Starts With Context — Not Your Prompt

The IDE does NOT send only your prompt.

It constructs a combined input:

Prompt + Code + Context + Tools
Enter fullscreen mode Exit fullscreen mode

Context includes:

  • Current file
  • Selected code
  • Nearby code
  • Related files (via search)
  • Available tools

2. Context Window — Why Results Differ

LLMs operate within a limited context window.

They:

  • only see what is provided
  • do not understand your entire project
  • do not know your intent beyond context

there is a reason why IDE's perform better with developers then so called non devs as the context differs in both cases.

3. No Explain Mode vs Write Mode

There is no mode switch.

The model predicts the best output format.

Explain:

Input: Explain this function
Output: Plain text
Enter fullscreen mode Exit fullscreen mode

Modify:

Input: Fix password validation
Output: Tool call
Enter fullscreen mode Exit fullscreen mode

4. Tools — The Execution Layer

AI does NOT modify files directly.
Tools are just predefined functions.
It calls tools like:

  • read_file
  • search_code
  • apply_patch
  • run_terminal

Example:

{
  "tool": "apply_patch",
  "args": {
    "path": "auth.js",
    "patch": "- if (password == null)\n+ if (!password || password.length < 6)"
  }
}
Enter fullscreen mode Exit fullscreen mode

5. How Code Is Edited

AI uses pattern-based editing.

Example:

- if (password == null)
+ if (!password || password.length < 6)
Enter fullscreen mode Exit fullscreen mode

Why not line numbers?

  • They change
  • Code shifts
  • Patterns are more stable

6. Feedback Loop 

Think → Act → Observe → Repeat
Enter fullscreen mode Exit fullscreen mode

7. Final Understanding

  • AI decides
  • Tools execute
  • Context drives everything

Conclusion

  • AI only knows what you show it
  • Better context → better output
  • Real skill = giving the right context
  • these IDE's are more than what i mentioned but this part is the one of the most interesting and important part to learn.

What’s Next

As we now understood abt tools the best thing to know next is MCP protocols
.

stayy tunedd!!!

Top comments (0)