Coding Agents: From a Single Terminal to Factories, Sandboxes, and Orchestrators

The rapid proliferation of coding agents capable of executing complex tasks over tens of minutes has left developers unable to keep up with multiple parallel sessions. This has created demand for dispatch programs that combine tmux (long-lived processes), git worktree (separate branches and directories for each agent), and isolation mechanisms such as Docker, gVisor, and Firecracker. In 2025, following the release of Anthropic's Claude Code, the first managers emerged: Claude Squad, Crystal, Vibe Kanban, Conductor, Superset, and Sculptor. By 2026, the genre had multiplied—dozens of projects like Agent of Empires, cmux, Pane, amux, and others implemented a similar formula. At the same time, older platforms evolved: Eclipse Theia (2017) integrated LLM and agent support through Theia AI (2024), while OpenHands, which started as a copycat agent of Devin, evolved into a control center for launching different agents. However, the market remains volatile: Vibe Kanban, which had garnered 27,000 stars, shut down in April 2026, and Crystal gave way to Nimbalyst. Conductor, on the other hand, raised $22 million in Series A funding. The key takeaway of the article is that the infrastructure for agents (terminals, isolation) was ready long before the emergence of powerful models, and it is precisely the combination of these components—not a new invention—that formed the foundation for the dispatchers.