ReAct Agent From Scratch

Overview

A ReAct (Reasoning + Acting) agent built in plain Python, without agent frameworks. It switches between Anthropic, OpenAI and local Ollama models through a single chat function, calls tools in a Thought → Action → Observation loop, and handles malformed model output and tool errors gracefully. It is built to understand what agent frameworks do under the hood, and is later rebuilt with LangChain for comparison.

Key features

• One interface, three providers: Anthropic, OpenAI and local Ollama models behind a single chat function, selected through environment variables.
• Thought → Action → Observation loop: the agent reasons, calls a tool, reads the result and repeats until it can answer.
• Defensive handling: malformed model output and tool errors are caught and handled instead of crashing the loop.
• Tested: the project uses pytest.

Tech

Python · uv · pytest · Anthropic API · OpenAI API · Ollama

Status

In progress. Source and updates are in the public repository.