open-agent-orchestrator 1.1.3

Control plane for AI agents

🔥 OpenAgentOrchestrator (OAO)

The Control Plane for AI Agents.

OpenAgentOrchestrator (OAO) is a Deterministic AI Execution Runtime (DAER) designed to bring infrastructure-grade governance, resilience, and observability to AI agents.

While most agent frameworks focus on building agents, OAO focuses on controlling them.

OAO acts as a control plane on top of existing AI frameworks, enabling safe, measurable, and scalable execution of AI agents.

---

📚 Technical Resources via OAO

We are building a library of technical content to help you engineer reliable agents.

📝 Technical Blogs

• Deterministic AI Execution: Why current agents fail in production and how to fix them.
• Why Agent Systems Need Governance: Implementing budgets, RBAC, and oversight.
• Replayable LLM Pipelines: Time-travel debugging for AI workflows.

🎮 Demos

• Failure Prevention: See StrictPolicy stopping runaway agents.
• Deterministic Replay: Resume crashed executions with zero state loss.

---

🛡️ Fault Tolerance & Persistence

💓 Robust Distributed Scheduler

• Crash Recovery: Automatically detects dead workers and re-queues their jobs.
• Heartbeats: Workers report liveness to prevent silent failures.
• Safe Claiming: Uses RPOPLPUSH to ensure zero job loss during assignment.
• Retries: Configurable exponential backoff for transient failures.

💾 Durable Event Sourcing

• State Reconstruction: Derives runtime state from immutable event logs for exactly-once correctness.
• Side-Effect Idempotency: Automated SHA-256 tool-call hashing prevents duplicate external actions during retries.
• Resume-on-Failure: Crashed workflows resume at the first incomplete step; completed work is skipped.
• Auditable History: Full execution trace stored in persistent storage (Redis or In-Memory).
• Time-Travel Debugging: Fork and replay past executions to reproduce bugs.

---

🚀 Why OAO?

Modern AI agent frameworks lack:

• ❌ Deterministic lifecycle control
• ❌ Strict policy enforcement
• ❌ Tool-level governance
• ❌ Execution observability
• ❌ Parallel scheduling control
• ❌ Infrastructure-grade architecture

OAO solves this.

---

🧠 Core Philosophy

OAO separates:

Agent Intelligence  ≠  Agent Governance

Frameworks build intelligence.
OAO governs execution.

Think of OAO as:

Kubernetes for AI Agents.

---

✨ Features

🧭 Deterministic Lifecycle Engine

Strict execution flow:

INIT → PLAN → EXECUTE → REVIEW → TERMINATE

No uncontrolled recursion.
No hidden state transitions.

---

🔐 Policy Enforcement

Built-in StrictPolicy enforces:

• Maximum execution steps
• Maximum token usage
• Maximum tool calls
• Maximum tool calls
• Execution timeouts

Violations trigger PolicyViolation events and halt execution.

Agents cannot bypass governance rules.

---

🔌 Adapter Architecture

Pluggable adapter system allows integration with external frameworks.

Currently supported:

• LangChain Adapter: With deep callback integration.
• LangGraph Adapter: Execute stateful graphs with managed telemetry.
• CrewAI Adapter: Kick off Crews or run individual Agents natively.
• AutoGen Adapter: Wrap conversation/initiations between ConversableAgents.
• Agno AI Adapter: Run Agno/Phidata agents with metric tracking.
• LlamaIndex Adapter: Wrap query engines and LlamaIndex agents.

Future roadmap:

• Enterprise custom adapters

Adapters are fully decoupled from orchestration core.

---

🔄 Async Execution Engine

Supports both:

• Synchronous execution (run)
• Asynchronous execution (run_async)

Ready for scalable, high-throughput workloads.

---

👥 Multi-Agent Orchestration

Run multiple agents under centralized governance:

• Independent lifecycle control
• Independent execution reports
• Controlled scheduling layer

---

🖥️ Observability Dashboard

Real-time visibility into the Deterministic AI Execution Runtime (DAER):

• Live Event Bridge: Stream execution events via WebSockets (/ws/events).
• Trace Timeline: Gantt-chart visualization of tool calls and step durations.
• Governance Watch: Real-time tracking of token consumption and budget violations.
• Scenario Simulation: Internal hooks for testing and fine-tuning agent behavior.

---

⚡ Parallel Agent Scheduler

Built-in concurrency management:

• Configurable max concurrency
• Async worker pool
• Safe task isolation
• Error containment

---

🌐 FastAPI Server (OAO as Service)

Expose OAO as an HTTP backend:

• Single-agent endpoint
• Multi-agent endpoint
• Swagger documentation
• Production-ready API layer

---

📊 Structured Execution Reports

Every execution generates:

• Unique execution ID
• Agent name
• Status (SUCCESS / FAILED)
• Total steps
• Token usage
• Tool usage
• Execution time
• State history
• Final output

Designed for observability and monitoring.

---

🎛 Event Hook System

OAO emits structured lifecycle events:

• STATE_ENTER
• TOOL_CALL
• POLICY_VIOLATION
• EXECUTION_COMPLETE

Hooks enable:

• Logging
• Metrics
• Monitoring
• External integrations

---

📦 Installation

Install from PyPI:

pip install open-agent-orchestrator

[!NOTE] By default, Orchestrator runs locally in memory out-of-the-box with zero Redis connection or package dependencies required.

Optional Dependencies

Install only the adapters you need to keep dependencies lightweight:

# Install with API server and specific framework support
pip install "open-agent-orchestrator[server,langchain,langgraph,crewai,autogen,agno,llamaindex]"

Or install locally:

pip install -e ".[all]"

---

⚡ Quick Start (Single Agent)

from oao import Orchestrator, StrictPolicy

class DummyAgent:
    def invoke(self, task):
        return {"output": f"Processed: {task}"}

policy = StrictPolicy(max_steps=5)

orch = Orchestrator(policy=policy)

report = orch.run(
    agent=DummyAgent(),
    task="Explain AI orchestration",
)

print(report.model_dump_json(indent=2))

---

⚡ Async Execution

import asyncio
from oao import Orchestrator

class DummyAgent:
    def invoke(self, task):
        return {"output": f"Processed: {task}"}

async def main():
    orch = Orchestrator()
    report = await orch.run_async(
        agent=DummyAgent(),
        task="Async execution demo"
    )
    print(report.model_dump_json(indent=2))

asyncio.run(main())

---

👥 Multi-Agent Example

import asyncio
from oao.runtime.multi_agent import MultiAgentOrchestrator

class DummyAgent:
    def __init__(self, name):
        self.name = name

    def invoke(self, task):
        return {"output": f"{self.name} processed: {task}"}

agents = {
    "researcher": DummyAgent("Researcher"),
    "critic": DummyAgent("Critic"),
}

async def main():
    multi = MultiAgentOrchestrator(max_concurrency=2)

    results = await multi.run_multi_async(
        agents=agents,
        task="Discuss AI governance"
    )

    for name, report in results.items():
        print(name, report.status)

asyncio.run(main())

---

🕸️ DAG Orchestration

Execute complex workflows with dependencies and automatic parallelism.

from oao.runtime.dag import TaskGraph, GraphExecutor, TaskNode

# Define graph
graph = TaskGraph()
graph.add_node(TaskNode("research", agent_researcher, "Research topic X"))
graph.add_node(TaskNode("draft", agent_writer, "Draft article", dependencies={"research"}))
graph.add_node(TaskNode("critique", agent_critic, "Critique draft", dependencies={"research"}))
graph.add_node(TaskNode("polisher", agent_polisher, "Improve draft", dependencies={"critique", "draft"}))

# Execute
executor = GraphExecutor(graph)
results = executor.execute("Write a blog post about AI")

Features:

• Topological Sorting: Ensures corect execution order.
• Cycle Detection: Prevents infinite loops.
• Parallel Execution: Independent branches run concurrently.
• Context Passing: Results flow from dependencies to dependents.

---

🌐 Run as API Service

Start server:

# Ensure server dependencies are installed
pip install "open-agent-orchestrator[server]"

uvicorn oao.server:app --reload

Open:

http://127.0.0.1:8000/docs

Available endpoints:

• POST /run
• POST /run-multi

---

📊 Observability (Metrics & Tracing)

OAO provides deep visibility into your agent fleets.

Prometheus Metrics

Exposed at /metrics:

• oao_executions_total: Execution counter (status, agent_type)
• oao_execution_duration_seconds: Histogram of execution time
• oao_active_agents: Gauge of concurrent agents
• oao_token_usage_total: Token consumption counter
• oao_queue_size: Distributed queue depth

OpenTelemetry Tracing

Full distributed tracing for workflows. Configure via OTEL_EXPORTER_OTLP_ENDPOINT.

• Root Spans: orchestrator.run, dag.execute
• Child Spans: oao.step.N, tool.execute, dag.schedule_task
• Context Propagation: Trace IDs flow across async tasks and Redis queues.

---

🔌 Enterprise Plugin System

Extend OAO without modifying core code. Built on a Secure Plugin Interface.

1. Create a Plugin (my_plugin.py)

Plugins must implement PluginInterface:

from oao.plugins.base import PluginInterface
from oao.policy.registry import PolicyRegistry

class MyPlugin(PluginInterface):
    @property
    def name(self): return "my_security_plugin"
    
    @property
    def version(self): return "1.0.0"

    def activate(self):
        # Register custom components safely
        PolicyRegistry.register("custom_policy", MyCustomPolicy)
        
    def deactivate(self):
        pass

2. Load the Plugin

from oao.plugins.loader import PluginLoader

# Verifies signature and version before loading
PluginLoader.load("path/to/my_plugin.py")

Supports custom:

• Policies (Governance)
• Schedulers (Execution strategy)
• Event Listeners (Logging/Tracing)
• Adapters (Framework support)

---

🏗 Architecture Overview

Client / CLI / Dashboard
            ↓
        FastAPI Server
            ↓
     OAO Orchestrator Core
            ↓
   Adapter → External Framework

• Lifecycle State Machine
• Policy Engine (Stop-Loss Governance)
• Adapter Registry (LangChain, LangGraph)
• Hash-Based Tool Interception Layer
• Append-Only Event Bus
• Execution Report Generator
• Parallel Scheduler
• Multi-Agent Coordinator

See the Detailed Architecture Guide for Mermaid diagrams and recovery flows.

---

🔒 Governance Model

OAO enforces:

• Deterministic state transitions
• Token budgeting
• Tool access limits
• Execution boundaries
• Timeout enforcement

Agents cannot override governance rules.

---

🧪 Project Structure

oao/
 ├── runtime/
 ├── adapters/
 ├── policy/
 ├── protocol/
 ├── server.py
 ├── cli.py

---

📈 Roadmap

• Deterministic lifecycle engine
• Strict policy enforcement
• Adapter abstraction
• Async execution engine
• Multi-agent orchestration
• Parallel scheduler
• FastAPI service
• Web dashboard
• Distributed scheduler (Redis)
• DAG-based orchestration
• Metrics exporter
• Enterprise plugin ecosystem
• Crash Recovery & Replay
• Tool Idempotency Wrapper
• Event-Sourced Determinism (DAER)
• OpenTelemetry Tracing
• LangGraph Support
• CrewAI Support
• AutoGen Support
• Agno AI (Phidata) Support
• LlamaIndex Support

---

🤝 Contributing

Contributions are welcome.

Guidelines:

• Maintain clean architecture principles
• Keep lifecycle deterministic
• Preserve adapter abstraction
• Add tests for new modules

---

📜 License

MIT License

---

🧠 Vision

OAO aims to become:

The Infrastructure Layer for AI Agents.

As AI agents become more autonomous, governance becomes essential.

OAO ensures agents remain:

• Observable
• Measurable
• Controllable
• Scalable
• Safe

---

⭐ Support

If you find OAO useful:

• Star the repository
• Contribute adapters
• Build plugins
• Share with the AI community

Let’s define the control plane for AI systems.