amux

The trait that decides it is unattended endurance. Recurring prompts on a schedule and self-pacing loops mean work continues while you sleep, and the watchdog compacts context and restarts sessions so a long run does not simply end at three in the morning. If your bottleneck is your own attention span, that is the feature.

It is also very new and very small, and it shows in the documentation. Pick it if overnight work is the point. Pick Crystal when two agents and your own eyes are enough.

The recovery mechanism is also a concealment mechanism. Automatic restart plus message replay means a session that dies repeatedly looks identical to one that is working, and the only evidence is a token bill that grows faster than the board does. Nothing documented describes a restart budget, an escalation or an alert.

What it does right is claiming. Cards are taken through a compare-and-swap, so two workers can never grab the same task, and that is a correctness property rather than a convention.

1. Distinguishing done from verified, and requiring different actors for each, imports a real quality practice into an agent fleet rather than trusting a single self-assessment. 2. The protocol's power depends on independence, and when both workers run the same model family a shared blind spot passes both stages.

2. Nothing published measures that. No agreement rate between checker and author, no false-pass figure, no comparison against a single-stage run. The idea is sound and the correlation is unmeasured.

Adding a no-selling restriction to a permissive licence is what a company does when it might want to sell hosting later and is not ready to say so. Behind it sits a business in a different category, which means this is funded by attention rather than by revenue, and attention is the scarcest thing on a small team's roadmap.

Moat: the clause, which stops resellers and attracts nobody. Likely path: it stays a demonstration of the parent's engineering, or it stops. Position: run it for personal work, and do not build a team practice on a project that has neither users paying nor contributors permitted to fork freely.

The licence restriction is irrelevant to internal use and does require a note from counsel, which is a week I would rather not spend. The larger problem is where the automation lives: scheduling happens inside this tool rather than in the pipeline system my team already operates, so I gain a second scheduler to watch and no central record.

Each of sixty developers would also need their own agent subscriptions behind it. Onboarding is an afternoon. Not yet: bring it back when it emits to something I already run.

Swapping the model in the middle of a session without dropping the conversation is the trick I have wanted for two years, and doing it in a single binary over an embedded database means there is no service mesh to stand up first. A local runtime is among the supported options, so a worker can run on my own hardware.

What costs it points is the licence restriction, which makes this source I may read and not genuinely mine, and there is no protocol interface, so my servers stay outside the fleet.