nac

The deciding trait is the dashboard. A long task is unbearable in a scrolling terminal, and opening a local web view from the project turns hours of activity into something you can glance at between other work. For anything that runs longer than your patience, being able to look without interrupting is the feature that decides whether you use it twice.

It is aimed at experiments and infrastructure work rather than tidy pull requests. Pick it for the long jobs. Pick a normal terminal agent for the short ones.

Privilege separation cuts both ways. Denying the planner any execution ability removes a class of accident and removes its only means of checking anything, so every decision it makes rests on a report it must accept at face value. A thread that overstates its progress does not produce an error; it produces a plan that proceeds as though the work were done.

What it does right is bounding the blame. When something goes wrong, the component that touched the disk is identifiable, because only one kind of component is allowed to.

1. This is the correct answer to the central problem in long-horizon agents. Raw transcripts grow without limit and summarisation applied late loses the wrong things; making the summary the interface means compression happens at a boundary designed for it. 2. The cost is stated implicitly: whatever the episode format omits is unrecoverable later, and the row does not define the format.

2. No evaluation of coherence over long runs is offered, which is precisely the property the architecture exists to deliver.

This is the strategically clearest project on the page. A firm whose business is models publishes a free harness whose first authentication option is its own, which puts the tool in the path of every task a developer runs. The permissive terms cost the company nothing and buy consideration at the moment a model is chosen.

Moat: the account relationship, not the code. Likely acquirer: none needed; this is the marketing budget. Likely path: maintained while it serves the model business. Position: adopt the harness, choose the model separately.

The use case is the right one. Experiments and infrastructure tasks are where unsupervised time goes missing, and a tool built for them could earn its place. Then the deployment story arrives: fetching and executing a script from a raw file host is not something my team permits, so packaging becomes ours.

There is no identity integration, no directory sync and no audit export, and it does not run unattended, which for a tool about long-running work is the surprise. Licence cost across sixty engineers is zero. Not yet.

The escape hatch is what makes this acceptable. A compatible key means I never have to hold an account with the company that wrote it, and permissive terms mean the fork survives whatever they decide next. Servers I already run attach over the protocol, so my existing capabilities carry across instead of being rebuilt.

The remaining gap is inference on my own hardware, which is not a supported target, so the one dependency I cannot remove is somebody's endpoint. Everything else here I can own outright.