Skip to content

A sophisticated recursive architecture that spawns and manages LLM-driven agents via a master-delegator model in a form of layers

License

Notifications You must be signed in to change notification settings

MajorTimberWolf/agent-orchestrator

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Autonomous Multi-agent Orchestration

Important

This branch consist of new proposal of architecture, Instead of parallel agents, I went with Layers of Agents approach. v1 will be available in old branch.

A sophisticated recursive architecture that spawns and manages LLM-driven agents via a master-delegator model in a form of layers (just like a neural layers), with task-specific prompt tuning, execution constraints, and multi-agent recursion without human intervention.

This uses a concept where a primary agent creates specialized sub-agents on demand. Each agent is dynamically generated with task-specific prompting and execution parameters, enabling efficient problem-solving without human intervention while not having a bias and having opinions from various agents instead of deciding on its own. Uses Ensemble technique.

Architecture

Here's how it works

architecture diagram

The system follows a recursive workflow:

  • Master Agent analyzes the input task and generates a comprehensive task breakdown with dependency structure.
  • For each subtask, the Master Agent creates a specialized Child Agent with custom fine-tuned prompts with dependencies.
  • Child Agents work on their assigned tasks and can create their own Sub-Child Agents when further specialization is needed
  • Child Agents are rearranged based on the dependency graph and co-ordinates with other agents while sharing context
  • Results flow back up the hierarchy for integration and final output
  • Results flow through layers of agents with context and intreprets a final output.

The key innovation is that agent creation and prompt engineering happen automatically at runtime in layers, with no predefined agent structures or human-designed prompts.

Caveats

As you see, this runs on recrusive layers method. building agents upon agents as it needs. This can result in building an agent layer network as shown below agent tree

The recursive nature of this system creates potential challenges:

  • Resource Management: Each additional agent consumes computational resources
  • Runtime Concerns: Deep agent layers can significantly impact completion time as it becomes sequential after a time

While the system offers powerful flexibility by giving control to the Master Agent, careful monitoring is recommended for resource-intensive tasks.

Demo

rao.mp4

Citations and Resources

  1. Emergence AI's 2025 Orchestrator Automatically creates agents and assembles multi-agent systems with minimal human intervention... continuously refining tools through recursive self-improvement

  2. ReDel's Recursive Systems (2024) Introduces systems where a root agent "decomposes tasks into subtasks then delegates to sub-agents" rather than using human-defined agent graphs.

  3. Beyond Better's Orchestrator (2025) Implements "sub-agents created with specialized capabilities for token efficiency and parallel processing" through dynamic task analysis.

License

This project is licensed under the GNU General Public License (GPL).

About

A sophisticated recursive architecture that spawns and manages LLM-driven agents via a master-delegator model in a form of layers

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%