Titan Orchestrator: A Distributed Agentic Workflow Orchestration Engine Built from Scratch

Titan Orchestrator is a zero-dependency distributed execution runtime built from scratch by independent developer Ram Narayanan. Its core goal is to bridge the gap between static DevOps pipelines and dynamic Agentic AI workflows, enabling unified orchestration of both through technologies like a custom DAG scheduler, the TITAN_PROTO binary protocol, and AOF persistence storage. The project is primarily positioned as an educational tool for learning distributed system principles, with production applications considered secondary.

Project Background

Titan was born from reflections on the complexity of modern orchestration systems, aiming to solve the problem of difficulty unifying static DevOps pipelines and dynamic Agentic AI workflows.

Design Philosophy

• Zero External Dependencies: The core engine is packaged as a single JAR file and can run without additional components.
• Education First: The README clearly states its goal is to help understand distributed system principles, rather than replacing production-grade solutions like Kubernetes or Temporal.

Three-Tier Capability Model

1. T1 Layer: Distributed task scheduler, suitable for batch processing, static DAGs, GPU/CPU routing, and other scenarios.
2. T2 Layer: Service orchestrator that supports long-running APIs and daemons, providing auto-restart and port management.
3. T3 Layer: Agentic runtime that supports self-mutating DAGs, LLM-driven Agents, multi-Agent pipelines, and HITL gating.

Custom Technology Stack

• TITAN_PROTO: A TCP-based fixed-header binary protocol that avoids JSON serialization overhead.
• Built-in DAG Scheduler: Processes complex dependencies between tasks.
• AOF Persistence: Enables crash recovery and state sharing through append-only logs.
• TitanStore: Optional distributed state storage that supports cross-node Agent state sharing.

Intelligent Routing and Scaling

• Capability tag routing (e.g., GPU, HIGH_MEM), affinity routing.
• Reactive auto-scaling: Spawns child processes when queues are saturated; idle nodes are retired after 45 seconds.
• Shortest connection distribution: Balances node loads.

Dynamic DAG Execution

Allows tasks to dynamically generate new tasks during runtime; Agents can autonomously decide the next step based on intermediate results, enabling intelligent workflows.

HITL Gating

• Supports pausing DAG execution at any checkpoint.
• Manual approval/rejection via dashboard, with a default timeout of 48 hours.
• SDK can automatically inject gating nodes.

Agent Runs Timeline

Groups all DAG stages of the same agent_run_id, clearly showing the complete lifecycle of multi-stage Agent iterations (PLAN→ITER→EVAL→SYNTH).

Visual Dashboard

• Orchestrator View: Real-time display of worker node status (capability tags, number of active jobs, etc.), with support for starting nodes via browser.
• DAG Pipeline View: Real-time rendering of dependency graphs; node colors update with status (PENDING→RUNNING→COMPLETED/FAILED), and stdout/stderr can be viewed.
• DAG Constructor: Drag-and-drop editor that supports configuring tasks, dependencies, and HITL gates; enables one-click deployment and generates Python SDK/YAML code.

Four Pipeline Definition Methods

Method	Best Scenario
YAML File	Reusable, version-controlled pipelines
Python SDK	Programmatic, runtime-dynamically adjustable pipelines
Visual Constructor	No-code drag-and-drop deployment
MCP (Natural Language)	Submit tasks via natural language using Claude/Cursor

MCP Integration

Built-in MCP server that supports describing requirements in natural language (e.g., researching three methods of distributed ML scheduling), automatically executing parallel jobs and synthesizing reports.

Deployment Modes

• Local Development: Single-machine run of Master+Worker+TitanStore+dashboard.
• Multi-cloud Deployment: Generate Master (2.3MB) and Worker (120KB) deployment packages via package_cloud.sh.
• Remote GPU Nodes: Local Master connects to cloud RunPod/VM as Worker via SSH tunnel.

Comparison with Existing Solutions

Feature	Titan	Kubernetes	Temporal
Number of Dependencies	Zero	Many	Many
Learning Curve	Steep but transparent	Steep	Medium
Agentic Support	Native	Requires additional layers	Limited
Dynamic DAG	Supported	Not supported	Not supported
HITL Gating	Native	Not supported	Not supported
Production Ready	Experimental	Mature	Mature

Summary

Titan represents a back-to-basics distributed system design approach, proving that a single developer can build a fully functional orchestration system. It is an excellent resource for learning distributed system principles, DAG scheduling, and Agentic workflows, with a clear architecture and rich documentation.

Outlook

• Current status: v1.0 experimental phase, Apache 2.0 license, single-master topology, process-level isolation.
• Future plans: v2 will support Raft consensus, Docker isolation, and mTLS security.