Field note
herdr describes itself as the runtime coding agents live on, and its scope is deliberately narrow: it does not wrap or replace Claude Code, Codex, Cursor, opencode or Grok, it owns their terminals. Panes keep running in a background server when the client closes or an SSH connection drops, and after a machine restart herdr restores the saved layout and can resume supported agent sessions, while being explicit that the original processes do not survive.
Two design choices make it more than a multiplexer. Every pane carries a state — working, blocked or idle — so when an agent stops and needs an answer, herdr says so instead of leaving a human to notice a quiet terminal. And agents themselves drive herdr through a CLI and socket API: they can spawn panes, prompt each other, and wait until another agent is genuinely blocked. That turns blocked into a coordination primitive rather than a UI hint, which is what parallel agent work actually needs.
Making the terminal a managed surface is the same consolidation as local coding agents becoming shared workspaces and the coding-agent workbench becoming its own product layer. The agent harness hub tracks that layer.
This is repository documentation: it describes behaviour and installation without reliability data or a list of which agent sessions actually resume. The condition to watch is whether resumable sessions are enumerated per harness, because a runtime that restores layout but silently loses one agent's state is harder to trust than one that admits it cannot resume.