Agent Orchestrator

English

中文

English

中文

Appearance

Agent OrchestratorHarness Engineering Control Plane

A local-first control plane that turns Claude Code, Codex, OpenCode, Gemini CLI, and shell tooling into a governed engineering harness for long-running software delivery loops.

Read the Vision
Get Started
GitHub
🤖

Harness Engineering Native

Built around a simple idea: humans define intent and constraints, while agents execute inside a reusable harness of workflows, skills, triggers, and guardrails.

🤝

Unified Agent Runtime

Claude Code, Codex, OpenCode, Gemini CLI, and other shell-based tools can be abstracted into the same runtime through capabilities and command templates.

📋

Declarative Control Plane

Workspaces, agents, workflows, triggers, secrets, and policies are defined as versioned YAML resources instead of scattered scripts and ad hoc prompts.

⚡

Long-Running Workflow Loops

Plan, implement, test, review, and fix can run as guarded, stateful loops with repeatable steps, DAG execution, dynamic step pools, and trigger-based task creation.

🔒

Policy and Recovery Rails

mTLS, RBAC, sandboxing, secret lifecycle management, loop guards, and persistent task state give agent execution stronger operational boundaries.

🌐

Observable by Default

Structured events, execution logs, task state, and machine-readable outputs make long-running agent workflows auditable and operable.

What This Project Is ​

Agent Orchestrator is not just an agent wrapper. It is a harness layer for software engineering systems.

OpenAI recently described this shift as Harness Engineering: engineers spend more effort designing environments, feedback loops, and control systems so agents can do reliable work. This project applies that idea to shell-native coding agents and local-first software delivery.

What This Enables ​

  • Institutionalize ad hoc agent usage into reusable manifests, skills, and workflow assets
  • Run durable plan -> implement -> test -> review -> fix loops instead of one-off agent calls
  • Keep one control plane while models, shells, and prompts change underneath

Quick Start ​

Run these commands directly, or let your AI coding agent handle it. You are booting a local control plane, not just invoking a CLI once.

bash
# Install
curl -fsSL https://raw.githubusercontent.com/c9r-io/orchestrator/main/install.sh | sh

# Or via Homebrew
brew install c9r-io/tap/orchestrator

# Or via Cargo
cargo install orchestrator-cli orchestratord
bash
# Start daemon and run your first task
orchestratord --foreground --workers 2 &
orchestrator init
orchestrator apply -f manifest.yaml
orchestrator task create --goal "My first QA run"
orchestrator task list