News · 2026-09-23
Orca: An Agent Development Environment for parallel coding agents
An open-source development environment built around isolated git worktrees for running and supervising multiple coding agents in parallel.
What it is
Orca is an open-source Agent Development Environment, or ADE, developed by Stably AI for running and supervising multiple coding agents in parallel. Instead of organizing work primarily around a single editor and a human writing code directly, it organizes work around isolated tasks.
Each task can receive its own git worktree, branch, files, agent terminals, editor state, and browser context. This allows agents such as Codex, Claude Code, OpenCode, and Pi to work on separate tasks — or alternative solutions to the same task — without modifying the same working tree.
Orca does not provide the underlying AI models. It sits above individual coding agents as the environment used to organize, observe, and coordinate the work they perform.
Core technical characteristics
- Worktree-native task isolation: each task can have its own branch, files, agent terminals, editor tabs, and browser tabs.
- Parallel agent execution: multiple CLI coding agents can run side by side, including separate solutions to the same task.
- Integrated terminals: terminal splitting, persistent scrollback, and multiple panes keep agent processes and development commands visible.
- Browser and Design Mode: rendered interface context, including HTML, CSS, and cropped screenshots, can be sent back into an agent prompt.
- Diff review and annotation: developers can inspect agent-generated diffs and turn line annotations into feedback.
- Remote execution: documented workflows include SSH worktrees, remote runtimes, cloud sandboxes, virtual machines, and local containers.
- Mobile companion: observe agents, receive completion notifications, and send follow-up instructions away from the desktop.
- GitHub and project integrations: pull requests, issues, and repository context can move into an isolated worktree.
- Orca CLI: worktree and environment actions can be addressed programmatically.
Why it matters
As coding agents become more capable, the problem becomes how to supervise several agents on a real repository without losing isolation, review discipline, or control over integration. Git worktrees provide a useful primitive: each agent operates in an independent checkout while still belonging to the same repository.
Orca turns that primitive into a broader workspace containing terminals, editors, browser state, diffs, and agent sessions. It addresses the environment around implementation, rather than replacing the coding agent itself.
Scaling AI-assisted development may depend less on a universally superior agent and more on reliable infrastructure around several specialized or competing agents.
Limitations and open questions
Worktree isolation is not security isolation. A worktree separates repository checkouts but does not inherently restrict operating-system access, network resources, environment variables, credentials, or external services. Stronger isolation needs containers, virtual machines, restricted users, or purpose-built sandboxes.
Parallel execution can increase model usage, compute requirements, and review workload. More candidate implementations do not automatically produce a better result; human review can become the bottleneck.
Independent, standardized evidence isolating the contribution of the orchestration environment itself remains limited. Orca is not evidence that parallel agents inherently outperform a well-configured single-agent workflow.
Supported agents, remote capabilities, integrations, and interfaces can change quickly. The official documentation and repository remain the reference for implementation decisions.