Agentic Blueprint Catalog

A collection of reusable patterns and blueprints for building agentic systems, with a focus on high-performance computing (HPC) and federated deployment scenarios.

Overview

This catalog provides production-ready examples demonstrating how to deploy AI agents across distributed computing environments using the Academy agent framework. The blueprints address common challenges in scientific computing workflows, including:

• Deploying agents to remote HPC systems via Globus Compute
• Orchestrating parallel task execution with Parsl
• Coordinating multiple agents across compute nodes
• Hierarchical work distribution patterns

Installation

pip install -e .

Requirements: Python 3.12+

Catalog Structure

agentic_blueprint_catalog/
├── agents/              # Reusable agent implementations
│   ├── pi_calculator.py # Monte Carlo Pi estimation agent
│   └── director.py      # MD simulation orchestration agent
├── federated/           # Remote deployment patterns
│   ├── remote_agent.py          # Client-launched remote agents
│   └── remote_spawning_agent.py # Self-spawning remote agents
├── hpc_hierarchical/    # HPC batch job patterns
│   ├── minimal.py       # Simplified hierarchical example
│   ├── complete.py      # Production Aurora deployment
│   └── aurora.yaml.j2   # Globus Compute endpoint config
└── model/               # LLM integration utilities
    └── model.py         # OpenAI-compatible model loader

Patterns

1. Federated Remote Agents

Deploy agents to HPC systems using Globus Compute, with the client orchestrating remote execution.

+--------+      +----HPC System-----+
| Client |----+-|--> Orchestrator   |
+--------+    +-|--> Simulator      |
              +-|--> Simulator      |
                +-------------------+

See federated/README.md

2. Remote Spawning Agents

Agents that dynamically spawn sub-agents on startup, enabling hierarchical orchestration where the client only interacts with a top-level coordinator.

See federated/README.md

3. HPC Hierarchical Tool Calling

Run multiple Director agents within a batch job, each managing a partition of compute nodes for parallel tool execution.

+---------------BATCH JOB-(4 Nodes)----------+
|  +-------Node0------+  +-------Node1------+|
|  | Director1 -----> PARSL POOL 1          ||
|  | Director2        |  |                  ||
|  +------------------+  +------------------+|
|  +-------Node2------+  +-------Node3------+|
|  |       ---------> PARSL POOL 2          ||
|  +------------------+  +------------------+|
+--------------------------------------------+

See hpc_hierarchical/README.md

Key Dependencies

Package	Purpose
academy-py	Agent framework for building and orchestrating agents
globus-compute-sdk	Remote execution on HPC systems
parsl	Parallel task execution and resource management
langchain	LLM integration and tool calling

Configuration

Copy and configure the environment file for LLM settings:

cp agents.env.example agents.env

For Globus Compute endpoints, see the configuration templates in hpc_hierarchical/.

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run linting
ruff check .
mypy .

License

MIT License - see LICENSE for details.

Authors

ModCon BASE Core Agentic Frameworks Team