Sherpa
AI-Augmented Thinking Companion
A model-driven Python framework for building intelligent multi-agent systems with Large Language Models. Sherpa combines hierarchical state machines with LLM reasoning to tackle complex, domain-specific tasks through structured workflows and agent collaboration.
Developed through multi-year collaboration between Aggregate Intellect and McGill University
Core Capabilities
Multi-Agent Orchestration
Build systems where specialized AI agents collaborate on complex tasks with shared memory and coordinated decision-making.
Flexible Memory Systems
Agent-specific belief states track execution history while shared memory pools enable cross-agent collaboration through vector stores and document repositories.
Policy-Driven Reasoning
Choose between rule-based deterministic logic or LLM-guided decisions. Automatic “fast-forward” optimization reduces API calls by ~50% when only one transition exists.
Production-Ready Features
Cost tracking, asynchronous execution, agent persistence across sessions, citation validation, and event-driven architecture.
Key Innovation
Model-Driven State Machines
Unlike other frameworks, Sherpa treats state machines as data structures, not implementation details. Design and modify complex workflows without changing code - enabling rapid experimentation and encoding domain best practices directly into your agent architecture.
Hierarchical Task Decomposition
Break down complex tasks into manageable substeps with composite states, guard conditions, and structured transitions based on proven domain workflows.
Peer-Reviewed Performance
Published research at MODELS 2025 demonstrates Sherpa improves task performance in 12 of 15 LLM-task combinations versus direct prompting, with particularly strong gains for structured tasks like code generation and domain modeling.
Architecture Highlights
Three-Layer Design
State Machines
Structure tasks into states, transitions, guards, and actions based on domain best practices.
Policies
Select transitions using rule-based logic or LLM reasoning depending on task requirements.
Belief System
Maintain execution context through trajectory history, action logs, and task-specific data.
The framework's decoupled design lets you experiment with different workflow configurations without touching implementation code.
Research & Open Source
Built on Academic Research
Sherpa is the outcome of a multi-year collaboration between Aggregate Intellect and McGill University. Our model-driven approach is backed by peer-reviewed research published at MODELS 2025, demonstrating measurable improvements in LLM task performance.
Open Source & Active Development
Contribute by reporting bugs, suggesting features, submitting improvements, or becoming a maintainer. Join our community in building better AI agent systems.
Future Directions
Human-Machine Teaming Lab
Sherpa is advancing through multidisciplinary collaboration with the University of Toronto - Rotman School of Business and the Alberta Machine Intelligence Institute (AMII) at the University of Alberta. This partnership is part of the Human-Machine Teaming Lab, a research initiative launched by Aggregate Intellect.
The lab addresses a fundamental question: How should organizations be designed to enable optimal human-machine collaboration?
Since early 2025, the team has been exploring this question by transforming a reinforcement learning gym into a knowledge work simulation environment. In this setting, multiple AI agents and human participants collaborate in simulated workplaces, executing complex tasks while navigating realistic information and communication constraints.
The research aims to identify emergent human-machine collaboration patterns and understand the organizational design factors that will shape the future of work.
To know more about this research or to propose collaborations, reach out to research@ai.science