Tau

The deciding trait is that you can read the whole thing in an evening. Every agent you use is a black box until you find one small enough to hold in your head, and after you have read this one the others stop being mysterious, because they are all doing roughly the same things with more code around them.

You are the wrong buyer if you want the most capable tool available today, because capability was not the goal. Pick it to learn and to work in a small way. Pick a fuller agent when you want volume.

The omissions are deliberate and the users will not be. There is no isolation layer and no commit boundary in the row, which is the right call for something written to be legible, and it means an agent that edits files and runs commands hands you no way back. Nothing in the packaging distinguishes a study object from a daily driver at the moment of installation.

What it does right is keep durable sessions on disk, so the record of a run survives the terminal that produced it.

1. The layering is the contribution. Provider differences are normalised into a neutral stream at the bottom, the middle owns messages, tools, events and the loop, and only the top layer knows it is a coding tool. Each layer can be read without the others, which is the definition of a successful decomposition.

2. That structure also makes the middle reusable for agents that are not about code. 3. No evaluation is published and none is needed, since the claim is pedagogical and the evidence is the source.

An actual company published this, which changes the survival question entirely. The tool costs nothing and is meant to cost nothing: 2,607 stars of goodwill from developers who now associate good taste with the publisher is the return, and that return is real even though no invoice exists.

Moat: the publisher's brand, which is the thing being spent here rather than built. Likely path: maintained while it serves that purpose, archived politely when it does not. Position: the safest bet in this cohort on continuity, and the least likely to grow into a product.

Procurement will move quickly on this, because the publisher is a company my legal team has heard of and the terms are standard. That is the whole easy part.

Operationally it gives me nothing: no console for sixty installations, no SSO, no audit log, no retention policy and no unattended mode, so it never becomes a step I can measure. The absence of any execution boundary means I would not put it near a machine with credentials on it. Approved with conditions: learning and local work only, on non-production checkouts.

The provider layer is the part I would use. It normalises vendors into one stream, which means adding an endpoint nobody thought of is a single adapter rather than a change threaded through the whole codebase. Three providers ship today, and the fourth is my afternoon.

There is no MCP in either direction, so my servers stay outside and the tool list is the built-in one. MIT, small, and honest about what it is, which means forking it is a reasonable plan rather than a threat.