HuggingFace released Tau, a coding agent

I'll pull the source article so the paragraphs stay factual and specific—no invented versions or figures.Hugging Face released **Tau**, a small open-source coding agent aimed at teaching how terminal coding agents are built rather than hiding the loop behind a black box. The project is documented at twotimespi.dev and lives in the huggingface/tau repository. Early discussion on Hacker News (AI Agents) was still thin — one point and no comments at the time of the linked thread.

**Tau** is split into three packages with a clear dependency direction. **tau_ai** turns provider responses into provider-neutral streams of tokens, tool requests, and thinking deltas. **tau_agent** is the reusable harness: messages, tools, transcript state, cancellation, queued prompts, and sessions. **tau_coding** wraps that harness with the coding environment — file and shell tools, durable sessions, skills, slash commands, and a Textual TUI. The design keeps the brain, the environment, and the UI apart so the harness does not depend on terminals, file paths, or rendering libraries.

For engineers and builders, the value is inspectability. **Tau** surfaces the questions most product agents skip: what an agent loop actually is, where tool calls come from, how the transcript grows, and how sessions survive process exit. Sessions are stored as inspectable JSONL under ~/.tau/sessions, with resume, branching, and export to JSONL or HTML. Local tools are typed for read, write, edit, and bash, and the stack includes context accounting, compaction, and thinking controls.

Market context is crowded with production coding agents (Claude Code, Codex, Cursor, and others) that optimize for finish-the-task UX. **Tau** sits closer to a curriculum and reference harness. It is inspired by **Pi** (pi.dev), which taught the same split between harness, environment, and UI; Tau is not a line-by-line port but a Python learning path over those ideas. Hugging Face is also positioning agents as first-class Hub citizens, so a readable open harness fits the company’s broader agent tooling story.

Install path is straightforward: `uv tool install tau-ai` (package name tau-ai). What to watch next is whether the community treats Tau as a forkable base for custom agents, how far provider adapters go beyond the documented options, and whether session export and branching become the practical way teams audit agent runs instead of trusting opaque chat history.