Agent Orchestrator
A local control layer for supervising parallel coding-agent sessions in isolated workspaces.
Per-dimension scores and reasoning
Evidence shows: README describes isolated workspaces, permission prompts (e.g., Pi reviewers with tools disabled), telemetry disclosure (PostHog, disableable), and reviewer permission modes (host-trusted vs user-approved). However, specific permission implementation details, user confirmation flows, data flow diagrams, sensitive data handling policies, dependency security audits, external impact scope, rollback mechanisms, and source attribution are missing. Deductions: insufficient evidence or only partial mentions.
Evidence shows: README and test files (e.g., install-check.sh, e2e-pod) indicate test coverage, but no test results or coverage rates are provided. Dependency availability is not explicit, failure messages are exemplified in test scripts (e.g., 'start did not fail with a clear error'), but overall consistency is not fully verified. Deductions: insufficient evidence for dependency availability, failure messages only partially demonstrated.
Evidence shows: README clearly identifies target users (developers) and scenarios (parallel coding agent management), supports multiple agents and platforms, and environment fit (macOS/Windows/Linux) is documented. However, trigger precision (e.g., specific conditions for automatic feedback loops) is not detailed, and capability boundaries (e.g., which operations are automated) are partially mentioned. Deductions: insufficient evidence for trigger precision.
Evidence shows: README has clear structure, detailed installation notes (desktop downloads, npm legacy), license is Apache-2.0, and documentation links are provided. However, naming stability (e.g., version naming) is not explicit, examples and FAQ are sparse, known limitations are not listed, versioning/changelog is missing, and maintenance responsibility is not clarified. Deductions: insufficient evidence for these aspects.
Evidence shows: README describes outputs (e.g., session states, feedback loops) and marginal value (parallel agent management), but cost-benefit (e.g., resource consumption, performance) is not mentioned. Deductions: insufficient evidence for cost-benefit.
Evidence shows: Claims in README (e.g., support for 25 agents) have some links, but no independent verification is provided. Facts and inferences are not clearly separated. Deductions: limited claim traceability and cross-source corroboration.
- Publisher identity is unverified; treat source with caution.
- Telemetry is enabled by default but can be disabled; be aware of privacy.
- Some reviewers (e.g., Agy, Continue) are marked experimental and lack OS isolation, posing security risks.
- Dependency security is not explicitly audited; verify yourself.
What does this agent do, and when should you use it?
Agent Orchestrator (AO) is a local Agent IDE and orchestration layer for running terminal coding agents in parallel. Its desktop app and CLI supervise projects, sessions, terminals, branches, pull requests, and feedback state while the selected agents perform the coding. Each session runs in its own Git worktree to separate concurrent changes. A local daemon watches session state, terminal activity, pull requests, CI, and review feedback, then routes relevant failures, comments, and merge conflicts back to the matching session. AO ships adapters for 23 worker harnesses and lists Claude Code, Codex, and opencode as separately configured reviewer harnesses.
A user adds a project in the desktop app or CLI and starts one or more sessions. AO creates an isolated git worktree for each session and launches the selected coding agent in that session's terminal runtime. Its local daemon watches session state, terminal activity, pull requests, CI, and review feedback; CI failures, review comments, and merge conflicts can be sent back to the relevant worker session. The desktop inspector exposes the selected session's terminal, pull-request state, review runs, and browser preview.
- An engineering lead running several independent changes in one repository and needing separate branches and workspaces for each coding session.
- A developer who needs to inspect a live Claude Code, Codex, or other terminal-agent session and send it follow-up instructions.
- A team that wants CI failures on a pull request routed back to the session that produced the change.
- A maintainer who runs reviewer agents and needs requested changes returned to the correct implementation session.
- A frontend developer who wants a local application preview beside the terminal for a session.
What are this agent's strengths and limitations?
- Uses a separate Git worktree per session, directly addressing file and branch isolation for concurrent work.
- Combines terminal access, session state, pull requests, CI, and review feedback in one desktop control surface.
- Ships adapters for 23 worker harnesses, with Claude Code, Codex, and opencode available as separately configured reviewers.
- Routes CI failures, review comments, and merge conflicts back to the relevant worker session automatically.
- Actual coding and authentication depend on separately installed supported terminal-agent CLIs; their setup is not detailed in the supplied material.
- Core operation relies on local Git worktrees, terminal runtimes, and a local daemon rather than a browser-only workflow.
- The npm CLI is frozen at 0.10.0, so new deployments are directed to the desktop application.
- The supplied material does not document offline behavior, CI/PR service authentication, permission controls, or retry behavior.
How do you install or deploy this agent?
Download the desktop build for macOS (Apple silicon or Intel), Windows, Linux AppImage, Debian/Ubuntu (.deb), or Fedora/RHEL (.rpm). Open Agent Orchestrator after installation and point it at the repository to manage; the desktop app runs the daemon. The legacy CLI remains available with: npm install -g @aoagents/ao followed by ao start, but the supplied documentation says npm is frozen at version 0.10.0 and recommends the desktop build for new setups. You also need a supported coding-agent CLI; its installation and authentication are not specified in the supplied material.
How do you use this agent?
Open the desktop app, add a target project, create one or more sessions, and select an available coding agent. AO creates a separate git worktree and starts an agent terminal for each session. Use the project and session views to monitor working, waiting, finished, or blocked state. In the selected session's inspector, review terminal output, PR state, review runs, or a local browser preview, and send follow-up instructions to that session.