Meta released a terminal-based coding agent aimed at people working across large codebases, alongside an update to the model powering it. The product competes directly with the equivalent tools from Anthropic and OpenAI.
Choosing the terminal as the product's home isn't an aesthetic call. It's where a developer already sits when working on a large system, and it avoids the context switch a separate window imposes. All three leading tools in the category converged on the same shape, which suggests the question is settled.
The feature presented as the differentiator is persistent background work: the agent keeps going on something slow while the person does other things. That shifts the model of use from question-and-answer to delegation, and brings along the perennial problem of knowing what it did while nobody was watching.
Meta was considered a straggler in this category, and the launch is the catch-up move. It's worth noting the coherence with the open weights announced days later: a tool that runs in your terminal on a model you can download is a different proposition from one that requires a cloud service.
For anyone running a development team, the point of attention is operational before it is technical: an agent working in the background needs audit logging and a bounded reach inside the repository, or code review stops knowing what it is reviewing.