A coding agent small enough to read in one sitting

What it does

Tau lives in your terminal and handles requests like “add tests” or “fix this stack trace” by reading files, editing code, and running shell commands. It keeps durable session history and streams its reasoning through either an interactive Textual TUI or a non-interactive print mode. Multiple model providers—including local endpoints—plug in via a neutral abstraction layer.

The interesting bit

The project is explicitly architected as a teaching tool. It splits into three readable layers: tau_ai for model abstraction, tau_agent for the portable brain (events, tools, loop), and tau_coding for the CLI and file-system glue. The core harness emits typed events and knows nothing about the TUI, so the same brain can drive a custom frontend or a library call without dragging in Textual or Rich.

Key highlights

• Durable append-only JSONL sessions under ~/.tau/sessions/ with resume and branching
• Context compaction, both manual and optional automatic, to manage token windows
• Project instructions auto-loaded from AGENTS.md, .tau/, and .agents/ directories
• Provider-neutral event stream renderable as Rich, plain text, JSON, or transcripts
• Usable as a library via AgentHarness, which accepts tools and models and yields events

Caveats

• Requires Python 3.12 or newer, which may limit where it runs
• Still under active development with a public roadmap of remaining work

Verdict

Ideal for developers who want to understand how coding agents are actually built rather than just using one. If you need a polished, zero-config IDE extension, this is not that.