Gemini coding agent · boot.dev
Built for the boot.dev AI agent course: a Gemini 2.5 Flash agent that plans with tools instead of guessing. It can inspect a project tree, read and write files, and execute Python — then loop on tool results until it has an answer.
Course project
Function calls until the job is done
The agent takes a user prompt from the CLI, talks to Gemini with a system instruction, and registers tools for listing directories, reading files, writing files, and running Python. Each model turn can request tools; results go back into the conversation for up to twenty-one iterations.
Capabilities
A small, honest tool surface
Working-directory injection keeps paths relative and sandboxed to the project. Verbose mode surfaces prompts and token usage. Unit tests cover each tool path — get files info, get file content, write file, run Python — so the agent’s hands stay predictable.
A bundled calculator package gives the agent a concrete codebase to inspect and fix during exercises.
Stack
Python · google-genai · uv
Python 3.12+, google-genai for the Gemini client, python-dotenv for API keys, and uv for lockfile-managed dependencies. Course work, kept public as a clean record of how tool-using agents are wired.
Problem
A coding assistant that only chatters cannot verify anything — it needs bounded tools against a real filesystem.
Approach
Gemini function calling with list/read/write/run tools and an iterative agent loop.
Outcome
A working CLI agent from the boot.dev course that can inspect and change a local Python project.
Technologies
- Python
- Gemini
- Function Calling
- CLI
- boot.dev