Agent Orchestrator: Plan, Run, and Supervise Coding Agent Fleets from One Place

Project Positioning: A Multi-Agent Orchestration Workbench for Agent-Driven Development

Agent Orchestrator (AO) is a local desktop workspace open-sourced by Untrivial.ai, tackling the emerging engineering problem of "multiple coding agents collaborating on a single project." One coding agent can handle a task, but running several across a project creates a different job: deciding what matters, splitting work cleanly, giving each agent the right context, preventing branch collisions, and following every change through review and merge. AO is built for that job.

The answer it offers is a "project-aware orchestrator": give every coding task its own agent, workspace, and feedback loop; track every worker, pull request, CI run, and review state in a live Kanban. Created in February 2026, roughly half a year later it has accumulated 10k+ stars and entered the GitHub Top 6k repositories — one of the most active open-source projects in the agent harness space today.

Core Abstraction: The Worker Session

The worker is AO's unit of execution: one task + one coding agent + one isolated workspace. Add a repository and create a worker session, choosing the coding agent, model, and interface that fit the task. For Git-backed work, AO gives the worker its own branch and worktree — the key design that prevents parallel agents from stepping on each other. The task, conversation, terminal, changed files, browser preview, pull request, CI, and review state stay attached to that session from start to finish.

Behind the desktop app, AO's local daemon watches agent activity and source-control state, assembling what would otherwise be a collection of disconnected terminals, branches, and browser tabs into a shared, live view of the project. The Kanban shows all sessions grouped by live status, so you can see at a glance what is moving, what is blocked, and where your attention will have the most impact.

The Orchestrator: Breaking Big Goals into Executable Plans

Beyond starting workers manually, AO provides a project-level orchestrator: describe a larger outcome, and the orchestrator shapes the plan using project context, creating workers on its own with the context and ownership they need. This forms a two-tier structure — the orchestrator handles "what to do and how to split it," workers handle "isolated execution," and humans handle "live supervision and decisions."

Agent Ecosystem: 26 Coding Agents, One Supervised Workflow

AO supports 26 coding agents through one supervised workflow, including Claude Code, Codex, Cursor, opencode, Aider, GitHub Copilot, Grok, Kimi, Pi, Amp, Auggie, Droid, Crush, Cline, Goose, Qwen, Continue, Devin, Kiro, Kilo Code, Vibe, Muse, Agy, Autohand, Kimchi, and Prime Agent. Agents keep their native strengths; AO supplies the project context, isolated execution, coordination, and operational view that make them work as a system.

You can use the interface that fits the moment: structured Chat or the agent's native terminal UI, while AO keeps task context, workspace state, and feedback in one place.

Engineering Highlights

• Pull requests and agent reviews: CI, mergeability, reviewer state, and interactive agent reviews beside the worker; requested changes return to the same owner;
• Agent-controllable browser: preview and inspect a worker's local app beside its interface; browser profiles are isolated per worker so parallel UI tasks don't share state;
• Isolation-first execution: every Git-backed worker gets its own branch and worktree; Scratch workers get AO-managed branchless directories;
• Live feedback loop: AO follows agent activity, pull requests, CI, review feedback, and merge conflicts, then reflects those facts on the Kanban.

Assessment

Agent Orchestrator hits precisely on the pain point of the current agent harness space: the bottleneck of coding productivity is shifting from "one agent completing one task" to "how to orchestrate a fleet of agents." The worker-session abstraction (task + agent + isolated workspace), the independent-branch worktree design, and the project-aware orchestrator together form a complete "agent-driven development" methodology. For teams already using multi-agent parallel development, this is one of the most mature open-source orchestration layers currently available. Code is released under Apache 2.0.