Fractal

The deciding trait is how narrow the tested ground is. The authors say it has been exercised on one operating system and make no promises about the others, which is an honest thing to write and a limiting one to read. On a Mac you are following a path somebody walked. Anywhere else you are cutting it.

What you get for that is a genuinely different idea about how an agent should be organised, running in your terminal today. Pick it for curiosity. Pick a boring terminal agent for work that has a deadline.

Three preconditions gate the first useful minute. The code path for running anything goes to a particular backend, that backend requires an authenticated session, and a default network policy must exist before the agent will act. Each of those is a separate thing that can be absent, stale or misconfigured, and the row records no fallback for any of them.

What it does right is routing execution through a container boundary at all, rather than running generated code on the host and calling the risk acceptable.

1. Recursion as the organising principle is a real architectural position: a task that exceeds what one pass can hold becomes sub-tasks of the same kind, and the composition rule is uniform rather than invented per situation. 2. That also means depth is the parameter which governs both capability and expense, and nothing in the row states a bound on it.

2. An architecture making this specific a claim invites measurement, and none is published. Recursion has a well-known failure mode, and the absence of a stated limit is the thing to check first.

The interest curve is running ahead of the product, which is a fine place to be if funding follows and a bad one if it does not. A company name and public discussion at this volume suggest somebody is building an audience deliberately, and audiences are the one asset in this category that can still be converted into a business later.

Moat: the runtime underneath, if it stays proprietary to them. Likely acquirer: a platform buying the team for the recursion work rather than this client. Position: track the company, not the binary.

A supplier describing its own software as alpha has told me everything procurement needs. I cannot put an alpha in front of sixty engineers, and I cannot write a change record around something whose authors decline to promise it works.

For completeness: no identity integration, no directory sync, no central audit record, and nothing that runs without a person watching it, so it never becomes a measurable step in delivery. The licence costs nothing and the model spend would be modest because usage would be low. Not yet, and not close.

The licence is the good part and it is not a small part: permissive terms on a genuinely novel runtime means the idea survives whatever the company decides about it later. Installation sitting alongside a Python toolchain I already use keeps the footprint small and reversible.

What I cannot do is the thing I most want to do here, which is run the recursion against models on my own hardware, because local weights are not a supported destination. There is no port for attaching servers I run either. The most interesting design on this page, behind somebody else's endpoint.