II-Agent

The deciding trait is what it hands back. Ask for a mobile or web application and you get one, not a patch series, which changes how you use it: you brief it like a contractor and review the output, rather than steering it file by file. For prototypes and internal tools that is a much shorter path than any editor agent.

The cost is control over the middle. You see less of the work while it happens. Pick it for throwaway builds and demos. Pick Manus when you want the same shape without running the stack.

The risk is the credential store. Eight named third-party services are wired in directly, and the same runtime that holds those tokens also drives a browser and executes commands, so a single poisoned page reaches mail, chat and file storage through one process. The blast radius is a person's whole working life, not a checkout, and nothing documented scopes one connector away from another.

What it does right: a plan stage exists before the build stage, so intent is inspectable and correctable while it is still cheap to change.

1. Output carries provenance. Research is delivered as a site with citations rather than as prose, which makes a claim checkable by a reader without rerunning anything. That is a better design choice than most deep-research tools make. 2. An interpreter executes generated code, so at least one class of error is caught by running it rather than by reading it.

2. No evaluation is published, and the documentation is a repository page, so the accuracy of those citations and the success rate of the builds are both unmeasured.

The structure is familiar: give away the runtime, host a convenience version, work out monetisation later. It works when the hosted tier eventually charges for something the self-hosted one cannot do. Nothing published says what that will be, and a research-flavoured vendor with no price list has not yet had to defend one.

Moat: the artefact quality, which is copyable, and 3,382 stars, which is attention rather than revenue. Likely path: a niche cloud service or an acqui-hire of the maintainers. Position: run it yourself, and treat the hosted app as a demo rather than a dependency.

Licensing for sixty engineers costs nothing and the deployment is a compose stack, which my platform team can stand up in a day and put behind our own gateway. That is genuinely better than most vendors in this category manage.

Then it stops. There is no unattended execution mode, so this is a thing people run at a desk rather than something scheduled, monitored and reported on. Provisioning and access control are whatever we build around it. Support is an issue tracker. Approved with conditions: one internal deployment, no production data, and a named owner on my side.

The model layer is the good part. The documented configuration accepts self-hosted endpoints alongside the usual hosted providers, so the weights can live wherever I decide, and every request is billed to a key I issued. Apache-2.0 means a fork stays legal and stays mine.

What annoys me is the protocol gap. It does not speak MCP, so the servers I already maintain are useless here and any tool I want has to be written into the project's own skill system. That is work I already did once. Still worth the trade for the licence.