Tau

Most agent frameworks bury their decision logic three abstraction layers deep, which means studying them tells you what they do but not how. Tau structures a working Python coding agent as an explicit three-package curriculum — tau_ai handles provider-neutral streaming, tau_agent is the reusable harness managing messages, tools, transcript state, cancellation, and sessions, and tau_coding adds the coding environment: file read/write/edit, bash execution, durable sessions, slash commands, and a Textual TUI. The vendor describes installing via uv tool install tau-ai and reading the code alongside running it as a real terminal agent.

The architectural principle the vendor emphasizes is boundary discipline: the agent harness must not depend on the terminal, file paths, or Rich rendering — those wrap the harness, they do not live inside it. This split between AgentHarness (the reusable brain), AgentSession (the coding environment), and the TUI (one possible frontend) is the core lesson Tau is built to teach. Sessions are stored as inspectable JSONL, which means you can export, branch, or replay a session without touching the agent internals.

Tau fits students, developers who want to understand agent internals before committing to a heavier framework, and contributors who want a modular base to extend. It does not fit teams that need a production-ready assistant, a hosted API, multi-user collaboration, or reliability at scale. At v0.1, the vendor frames the roadmap as a phase-by-phase build — each release adds a layer with notes on what was added and why, which makes it a moving curriculum rather than a stable dependency.