AI Agent in Python: Build a Coding and Search Agent Loop

A new tutorial shows how to build an AI agent in Python in ~131 lines. The core pattern is the same for different use cases: connect an LLM, add tools, run an agentic loop, then run a conversational loop. The coding agent uses four tools—read, write, edit, and bash—so it can organize files, wrangle data, manage media, and run code via the shell. The bash tool turns a coding agent into a general-purpose “computer-using agent,” but it is flagged as dangerous and should run in a sandbox/container/VM. The post also builds a search agent using Gemini plus a web_search tool backed by Exa, to ensure current web information instead of relying on model training data. For multi-step tasks (e.g., comparing two items), the AI agent keeps calling tools until it has enough evidence, with message history preserved across turns. Key message: the “magic” is not complex algorithms, but the AI agent loop and well-designed tool interfaces that safely return results to the model for iterative improvement. Crypto-trader relevance: this is mainly developer infrastructure news for AI agents, not a protocol change, but it can accelerate tooling for automated research and code-driven workflows.