Hive

The deciding trait is that the coordinator is an agent rather than a script. You describe the job once and something with judgement decides how to break it up, instead of you playing project manager across six terminal windows and forgetting which one was doing the migration. When it works, the supervision cost genuinely drops.

When it does not, you are debugging a delegation rather than a program, which is a stranger afternoon than it sounds. Pick it for work with obvious parallel parts. Pick one focused session for anything you could have finished yourself in an hour.

Delegating the staffing decision delegates the budget. A coordinator that can hire and dismiss workers as it sees fit will size the team to the task as it understands the task, and its understanding is exactly the thing under test. Nothing documented caps how many processes a single request may produce. The failure is not a crash; it is six sessions of an expensive model all reading the same repository.

What it does right is make the workers visible. They are real terminal processes you can watch rather than hidden calls in a transcript.

1. Using a text file as the shared blackboard is a deliberately low-technology choice, and it has one large virtue: the coordination state is readable by a human without any tooling, which is more than most orchestrators offer. 2. Injecting a command into each shell means delegation travels the same channel as the work, so no separate transport has to be kept alive.

2. The cost is that concurrent writers share one document with no described locking discipline, and the phase fan-out built on top is labelled experimental by the project itself.

Every expensive component in this system belongs to somebody else. The agents doing the work are other companies' products, authenticated with their logins and billed on their meters, and what is owned here is the arrangement between them. That is a genuine contribution and a thin asset, because the vendors underneath are all building their own versions of it.

Moat: none against the vendors, some against peers while the ergonomics stay ahead. Likely acquirer: whichever agent vendor decides orchestration should be first-party. Position: useful now, and do not architect a team around it.

The default I appreciate is the one that does nothing: remote access from a paired phone stays disabled until a person enables it, which is the correct posture and the opposite of what most products ship. That is one fewer conversation with my security team.

The platform rating is the problem. Best-effort support with a partial test suite is not a commitment I can put in front of the engineers who work on it, and it means sixty seats really means the subset on other machines. No console, no directory login, no audit trail, nothing unattended. Not yet.

I can read the code and I cannot own it, and the version where I could is behind me. A business-source licence with a conversion date years out is a bet that the project matters enough for people to accept the terms in the meantime, and it removes the one guarantee I actually care about: that a fork stays legal on the day the maintainer loses interest.

Everything else is fine. One global package install, agents authenticate with the logins already on my machine, and no protocol layer of its own to configure or disable.