Tau (τ) — Educational Coding Agent

2026-07-03 · source · source ↗ · raw capture · wiki page

Tau is an open-source (MIT) terminal coding agent by Alejandro Ao (GitHub: alejandro-ao), published on PyPI as tau-ai. It's designed to be both a working tool and an educational resource — "a small Python coding agent you read like a textbook." The project is inspired by Pi (Mario Zechner's minimal coding agent) but is an original implementation.

Three-Layer Architecture

tau_ai — Provider adapters translating OpenAI, Anthropic, Hugging Face, OpenRouter, and local endpoints into Tau's provider-neutral event stream.
tau_agent — The reusable harness: messages, tools, events, agent loop, session primitives. Provider- and UI-agnostic.
tau_coding — The coding environment: file/shell tools, sessions, skills, slash commands, and a Textual-based TUI.

Dependency direction: tau_coding → tau_agent → tau_ai.

Design Principles

Small layers beat magic — One job per package, each studyable independently.
Events make agents teachable — The agent emits a typed event stream you can render, test, and export.
Real enough to matter — Educational, not a toy. Runs as a real terminal agent.
Docs follow implementation — Built phase by phase with notes on each addition.
Tools are ordinary typed functions — Schema + async executor, no complex framework.
Sessions are durable and inspectable — Append-only JSONL under ~/.tau/sessions/ with resume and branching.

Features

Interactive TUI (tau) and non-interactive print mode (tau -p "prompt")
Four built-in tools: read, write, edit, bash
Slash commands: /login, /model, session management, compaction, export, theme switching
Project instructions from AGENTS.md, .tau/, .agents/
User skills and prompt templates
Context accounting with manual/automatic compaction
Multiple provider support normalized through tau_ai
Programmatic use: import AgentHarness, iterate over events from harness.prompt()

Educational Value

Tau demonstrates: provider-neutral streaming, agent loops with tool calling, typed local tools, durable sessions, session resume/branching, JSONL/HTML export, project instructions, skills, prompt templates, slash commands, context accounting, compaction, thinking controls, and UI adapter boundaries.

The project has 438 commits, primarily Python (93.7%), with documentation in Astro/HTML/JS/CSS. The core lesson: "Separate the brain, the environment, and the face."

Repository

https://github.com/alejandro-ao/tau — MIT License, 438 commits.