Kiwi Code DocsHow Kiwi Code Works
The agentic loop — how Kiwi Code thinks, acts, and iterates to complete your tasks.
How Kiwi Code Works
When you send a message to Kiwi Code, the AI doesn't just generate text — it enters an agentic loop where it thinks, uses tools, observes results, and repeats until your task is done.
The Agentic Loop
Every interaction follows this cycle:
You send a message
│
▼
┌─────────────┐
│ AI Thinks │ ← Analyzes your request + context
└──────┬───────┘
│
▼
┌─────────────┐ ┌──────────────────┐
│ Tool Call? │────→│ Execute Tool │
│ (yes/no) │ │ (on your machine │
└──────┬───────┘ │ or in cloud) │
│ └────────┬──────────┘
│ │
│ ┌───────▼──────────┐
│ │ Observe Result │
│ │ (feed back to AI)│
│ └───────┬──────────┘
│ │
│ ┌──────▼──────┐
│ │ Loop Again │──→ back to "AI Thinks"
│ └─────────────┘
│
▼ (no more tool calls)
┌─────────────┐
│ Response │ ← Final answer streamed to you
└─────────────┘The AI keeps looping — thinking, calling tools, observing results — until it has enough information to give you a complete answer or has finished making all the changes you asked for.
A Real Example
When you ask "Add input validation to the User model", here's what happens:
Think
The AI reads your message and decides it needs to see the current code first.
Act — Read the file
Tool call: read file src/models/user.py
The runtime reads the file on your machine and sends the contents back.
Observe
The AI sees the current User model code and plans the changes.
Act — Edit the file
Tool call: write file src/models/user.py with the updated code including validation.
Act — Run tests
Tool call: run command pytest tests/test_user.py
The runtime executes the tests on your machine.
Observe
Tests pass. The AI is satisfied the changes work.
Respond
The AI explains what it changed and why. Loop ends.
This entire sequence — 3 tool calls, multiple think-observe cycles — happens automatically from a single message.
Tools the AI Can Use
Each iteration of the loop, the AI can call any of these tools on your machine through the runtime:
| Tool | What it does | |
|---|---|---|
| Run commands | Execute any shell command — build, test, lint, install packages | |
| Read files | Read file contents, search within files, list directories | |
| Write files | Create, edit, or delete files — apply diffs, find-and-replace | |
| Search code | Grep across your codebase with pattern matching | |
| Web search | Search the web for documentation, examples, or solutions | |
| Fetch URLs | Read web pages, API docs, or online resources |
Tools are automatic
You don't need to tell the AI which tools to use. It decides based on your request — ask it to "fix the tests" and it will read the test file, run the tests, read the errors, edit the code, and re-run until they pass.
Loop Limits
The AI doesn't loop forever. Built-in limits keep things under control:
| Limit | Value |
|---|---|
| Max iterations | 20 per message (50 for coding tasks) |
| Command timeout | Up to 600 seconds per command |
| Response timeout | 15 minutes per AI call |
If the AI hits the iteration limit, it stops and tells you what it accomplished and what's left to do. You can continue by sending another message.
Real-Time Updates
While the loop runs, you see live updates in the TUI:
- "Thinking" — the AI is processing
- "Executing" — a tool is running on your machine
- "Streaming" — the AI's response is being generated
Open the runtime logs (Ctrl+O) to see exactly which commands are being executed.
How It Differs by Agent
All agents use the same agentic loop, but with different AI models:
- Deep Thinker agents spend more time in the "Think" phase — better for hard problems
- Flash agents think faster — better for simple tasks
- Standard agents balance speed and depth
The tools and loop mechanics are identical across all agents.