fx

You will like this if you think of agents as components rather than applications. It is one compiled binary meant to be dropped inside a larger system, which is the deciding trait, because everything else in this category assumes it is the thing you sit in front of and this one assumes it is a part.

Pick it if you are building something that needs an agent inside it. Pick OpenCode when you want a terminal agent that is stable, complete and someone else's problem to maintain.

A month of public existence is not enough time for the failure reports that make a tool trustworthy to have been filed, let alone fixed, and the project says preview itself rather than leaving you to discover it. Against that youth sits command execution on your working tree with no isolation on this row, which is the combination that produces the stories people tell afterwards.

What it does right: independent work is delegated to subagents with their own contexts, so one long task does not poison the context of everything after it.

1. Keeping output closer to a shell than a full-screen application is an architectural decision rather than a stylistic one: text that scrolls can be piped, logged and diffed, which a redrawing interface cannot. 2. Reading a configuration format established by another vendor means a user's tool servers move between agents without translation, which is the correct answer to a fragmenting ecosystem. 3. Skills provide the extension surface.

No benchmark is published for a month-old project. The observation: adopting a rival's config file is a small act of humility that saves everyone work.

The label matters commercially. A labs release carries no roadmap commitment and no support obligation, which is exactly the right structure for an experiment and exactly the wrong one for a dependency. Underneath, the sign-in options route inference through the parent's own gateway, so a permissively licensed binary becomes a meter on the parent's revenue line while costing them almost nothing to publish.

Likely path: graduation into a supported product, or quiet retirement, and labs projects go both ways. Position: watch it, do not build on it, and read the graduation announcement as the actual product launch.

One sentence on the demo: it printed like a shell and behaved like one. Everything after that is a gap. Installation is a shell script per developer, so there is no console, no policy layer and no directory integration, and inference is billed against a gateway account or an individual's consumer subscription, neither of which produces a per-team number I can forecast. Windows engineers are excluded. Nothing runs unattended.

Onboarding is under an hour for anyone comfortable in a terminal. Not yet.

Trust being explicit is the detail I keep wanting elsewhere. A server sits inert until I approve it, so adding one is a decision rather than a side effect of editing a file, and that is the right default for anything that can call out of my machine. The licence is permissive and the source is a systems language, so a fork compiles to something I can ship.

The complaint is authentication. Getting in means the parent's gateway or one of two consumer subscription logins, and no local model runtime is documented, so my own hardware is not on the list. Small binary, short leash.