Skip to content
kaustubh

basic-ai-agent

An AI coding agent, built in Python following boot.dev's course. Next, rebuilding it in TypeScript with Effect, and giving it web search.

What it is

A small command-line coding agent. Give it a task in plain words and it reads, writes and runs files inside one working folder until the task is done.

Why I built it

I use AI agents every day and wanted to know what actually happens between the prompt and the answer. The course got me a working loop; the rest is me pushing it further.

How it works

Flowchart of the agent loop: a task goes to the model, which either calls list_files, read_file, write_file, or run_python_file and receives the result back, or stops and answers “done”.

  • The model gets the task plus a list of tools: list files, read a file, write a file, run a Python file.
  • Each turn it either calls a tool or answers; tool results go back into the conversation.
  • The loop stops when the model answers without calling a tool, or after 20 turns.

What I learned

  • An agent is mostly a loop around tool calls, not magic.
  • Tool descriptions matter as much as the prompt.

My vision (or goal)

  • Rebuild it in TypeScript with Effect: every way a step can fail (a missing file, a model timeout, a rate limit) becomes a typed error I have to handle, model calls get retries with backoff, and the tools become swappable services so the loop runs against a fake model in tests.
  • Add a web-search tool, so the agent can look things up instead of only touching files in its folder.
  • Why Effect: I want to learn how to write async code where failure is part of the design, not an afterthought.