HuggingFace released Tau, a coding agent

I'll pull the source article so the paragraphs stick to real details about Tau—no invented versions, dates, or figures.**Hugging Face** released **Tau**, a coding agent presented as both a usable terminal tool and a teachable reference implementation. Coverage appeared on HN AI Agents via the project site at twotimespi.dev, with the linked Hacker News thread at low early visibility (1 point, 0 comments). Tau is framed as a small Python agent you can read like a textbook while it runs: stream model output, call tools, manage sessions, and grow into a terminal UI.

Technically, Tau is split into three layers with a one-way dependency path: **tau_ai** (provider adapters that turn model responses into provider-neutral streams of tokens, tool requests, and thinking deltas), **tau_agent** (the reusable harness for messages, tools, transcript state, cancellation, queued prompts, and sessions), and **tau_coding** (the coding environment—files, shell, durable sessions, skills, slash commands, and a Textual TUI). The design lesson is the boundary: AgentHarness stays free of terminal, file-path, and Rich concerns; AgentSession is the coding environment; TUI is one frontend. Sessions live under inspectable JSONL history (with resume, branching, and HTML export), and local tools cover read, write, edit, and bash.

For engineers and builders, the value is visibility into parts most coding agents hide. Tau answers concrete questions tutorials skip: what an agent loop is, where tool calls come from, how the transcript grows, and how sessions survive process exit. You can run it as a real terminal agent—requests like explain this repo, add tests, or fix this stack trace—while studying the packages that implement provider streaming, the loop, and the coding app in isolation. Event streams, not buried callbacks, are the contract between layers, which makes the system easier to render, test, and export.

In market terms, Tau sits next to opaque commercial coding agents and closed harnesses by competing on legibility, not feature race. It is explicitly inspired by **Pi** (pi.dev): keep harness, environment, and UI apart. Tau is not a line-by-line port; it is an educational Python take on the same split. Install is via `uv tool install tau-ai`. Provider setup includes Hugging Face Inference Providers and other model backends, so builders can swap models without rewriting the agent loop.

Practical takeaway: if you are building or evaluating coding agents, treat Tau as a map—events first, then the loop, then harness, tools, and UI—not as a black-box product to adopt blindly. Watch how its package boundaries hold as skills, slash commands, context accounting, compaction, and thinking controls land; those are the seams where real agents either stay modular or collapse into a monolith. For a first pass, install it, run a session against a real repo, then read **tau_ai** → **tau_agent** → **tau_coding** in that order.