Designing the tiled session grid
Running one agent at a time is simple, but it leaves a lot of attention unused. The grid is Blackcrab's attempt to make parallel agent work legible without turning it into a control room.
Each tile is a work unit
The grid works because each tile has a clear owner: one session, one transcript, one composer, and one working directory. That keeps parallel work from collapsing into one long feed.
A tile should answer the practical questions immediately. What is this session doing? Which repo is it in? Is it still streaming? How much context is left? What changed since I last looked?
Density without dashboard noise
Developer tools can become decorative dashboards quickly. The grid tries to stay closer to a terminal multiplexer: dense, predictable, and built around repeated actions instead of one-off presentation.
That means stable tile sizes, keyboard focus, compact status, and enough visual contrast to scan six sessions without needing a legend.
- Command palette for cross-session navigation.
- Numbered shortcuts for direct tile focus.
- Per-session state instead of global mode switches.
- Compact status text that does not resize the layout.
The open question
The hard part is deciding how much automation belongs in the grid itself. Blackcrab should help notice stuck sessions and expensive runs, but it should not pretend to know which thread deserves your attention.
For now, the product favors explicit control. The app makes many sessions visible, but the developer still chooses when to split work, when to merge it, and when to stop an agent before it wanders.