Arkain

The deciding trait is the review gate. Every change lands in a diff viewer and every command waits for an accept or a reject, so the agent works in front of you rather than behind you. If you have watched an autonomous tool rewrite four files while you were reading the first one, that pacing is the whole difference.

The cost of that pacing is that you are always there, which makes this a companion and not a delegate. Pick it if you want a browser tab and no setup. Pick a local terminal agent if you want the machine and the bill to be yours.

The gap is the return trip. Code arrives from a repository, gets edited, and then the documentation stops. Nothing describes a commit, a branch or a pull request produced by the agent, which means every accepted change still needs a human to carry it home. For a product whose pitch is autonomy inside the workspace, that is the least autonomous part of the loop.

What it does right is the permission model. Agent permissions are configurable and each command asks before it runs, so nothing here pretends the accept step is optional.

1. Exploration is asserted rather than specified. Whether context is assembled by search, by an index, or by reading directories in sequence changes both the cost and the failure profile, and the documentation does not say which. 2. No published evaluation accompanies the capability, so the claim is a description of intent.

2. One design choice is genuinely principled: execution and editing happen in the same environment the code will run in, which collapses the usual gap between what an agent verified and what the developer will observe afterwards.

The agent is not the business. Compute is, billed at $0.272 an hour per container plus traffic and storage, and the model sits on top as the reason you leave the tab open. That is a durable shape: the meter runs whether or not the agent is any good, and an established vendor with an existing customer base is underwriting it.

Moat: an existing Korean developer install base and switching cost measured in configured environments. Likely path: the agent stays a feature of the container business rather than becoming its own line. Position: sound vendor, narrow geography.

The subscription is the small number. Eight hundred and forty dollars a month across sixty engineers is approvable; the metered compute underneath it is the line finance will ask about in month three, and I cannot forecast it from this page.

Procurement is where it stops. There is no directory integration, no provisioning, no audit trail of what an agent executed in a container holding our source, and no retention statement I can attach to a questionnaire. Nothing runs in a pipeline, so it never shows up as a measurable step. Not yet.

There is nothing to fork and nothing to read. The licence is proprietary, I cannot supply my own key, and running a local model is not on offer, so every token goes through somebody else's account at somebody else's chosen provider. When it behaves badly I file a ticket, which is not a relationship I enjoy.

The one genuine extension point is MCP, and I will grant that it is the right one: my own servers become tools inside their container. That is real, and it is the only part of this I own.