Agents and the Genesis Mission

Background and technology recommendations

These slides provide some background and initial recommendations concerning appropriate agentic technologies for use in Genesis Mission projects.

​

They are very much a work in progress.

​

Comments and questions to foster@anl.gov

​

​

Why agents?

The Genesis Mission seeks to create autonomous, high-throughput, continuously operating scientific systems that integrate computation, experiment, data, and models across the national research infrastructure

Agents have promise as a source of the operational intelligence needed to coordinate, adapt, and accelerate discovery at scale, by:

• Enabling autonomous scientific operation
• Connecting diverse systems into a coherent whole
• Enabling adaptive experimentation and discovery
• Providing scalability across thousands of instruments and workflows
• Enabling mission-level autonomy and scientific governance
• Providing appropriate abstractions for human–AI collaboration

​

​

​

Questions that we address here

• What is an agent, anyway?
• Appropriate agent technologies for use in GM applications
• Early examples of agents applied to GM problems (pending)
• How AmSC and ModCon are working to facilitate use of agents in GM applications

​

These slides will be updated; check back soon

​

Agents background

What is an “agent”?

An agent is a persistent, stateful process that acts on behalf of a user or system. An agent may:

​

• Observe inputs or events
• Plan (decide on) actions using a policy (rules or LLM)
• Act: Execute tools or call other agents
• Learn: Update state to adapt over time

​

We can think of an agent as a scientific assistant that can reason, act, and coordinate on our behalf

Two common patterns: Workflows & agent networks

• Agents make only internal decisions: Which tool to run, response to give
• Fixed sequence of steps
• Deterministic, reproducible
• No comms with other agents

• Agents also make external decisions: Which agent to contact or delegate to
• Dynamic, adaptive collaboration
• Enables negotiation, coordination, and distributed workflows

​

​

Workflows / chains

Agent Networks

Other distinctions

• An AI agent uses an LLM or reasoning model to make decisions; a dumb agent uses simple decision rules
• A learning agent dynamically updated its state over time
• An agent may be persistent, e.g., to process jobs submitted to a queue or new hypotheses or observations, or short-lived
• An agent may or may not be tool calling
• A conversational agent engages in dialog with humans or other agents

​

In summary

Scientific agents span a spectrum:

• From simple tool wrappers
• To LLM-driven planners
• To communicating agent networks
• To continuously running scientific systems

​

No single technology covers all cases. Choose based on autonomy, structure, scale, and environment.

​

​

Agent frameworks

There are many agent frameworks

LangGraph

Microsoft Agent Framework (prev. AutoGen)

Academy (UChicago)

CrewAI

Goose (Linux Foundation)

Google agent SDK

SmolAgents (Hugging Face)

LlamaIndex

…

We recommend: see below

Agent frameworks simplify the creation of multi-step (LLM) workflows

They do so by managing prompts, context, state, memory, and tool invocation to support reliable reasoning and action

Different frameworks focus on different types of activities

None apart from Academy address explicitly the unique needs of DOE science (e.g., federation, HPC)

There are benefits to focusing on a few frameworks in GM

Our recommendations (a work in progress!)

• LangGraph offers structured, reproducible workflows for LLM-driven reasoning and tool execution. Good for managing interactions with LLMs and for implementing structured or auditable reasoning pipelines.
• Microsoft Agent Framework (MAF) supports flexible multi-agent coordination patterns and conversational planning. Good for multi-agent coordination, committee-based reasoning, or adaptive planning strategies involving interaction among several agents.
• Academy provides persistent, secure, and scalable execution across HPC systems, instruments, and data services. Good for agents that must run continuously or securely on HPC systems, laboratory robots, data platforms, or other parts of federated DOE infrastructure.

​

These roles are complementary rather than competitive. For example, teams may use LangGraph to manage execution of federated Academy agents.

LangGraph: Structured LLM Workflows

• A framework for building deterministic, controllable, multi-step LLM workflows using directed graphs, providing:
• Graph-based orchestration of actions and tools
• Deterministic control flow with optional branching
• Local, per-run state management
• Easy integration of LLM reasoning with scientific tools
• Good for:
• Multi-step tool use
• Retrieval + reasoning pipelines
• Scientific analysis chains
• Workflows that must be predictable and auditable
• Provides structure around LLMs, turning complex reasoning tasks into transparent, reproducible workflows

https://www.langchain.com/langgraph

(and predecessor LangChain)

Microsoft AutoGen: Conversational Multi-Agent Systems

(also known as Microsoft Agent Framework)

• A framework for building agents that collaborate through messaging, often using LLMs to drive dialogue and delegation, providing:
• Multi-agent communication out of the box
• Conversational task planning
• Dynamic, emergent coordination behaviors
• Simple abstractions for expert agents, delegates, critics, etc.
• Fast prototyping of social or committee-based agents
• Good for:
• Agents that send messages to each other
• Expert-group debate or consensus formation
• Brainstorming, planning, critique workflows
• Early exploration of multi-agent behaviors
• Enables flexible, dynamic collaboration among agents; ideal for early experimentation or decision-making workflows

​

​

https://github.com/microsoft/agent-framework

Academy: Persistent Scientific Agents at Scale

• A platform for persistent, registered, secure scientific agents that operate across HPC systems, labs, instruments, and data services, providing:
• Long-lived, identity-bearing agents
• Integration with federated cyberinfrastructure (HPC, instruments, storage)
• Secure execution, authentication, and routing
• Event-driven and continuous operation
• Internal tool use + external agent delegation
• Good for:
• Autonomous laboratories and experiment loops
• Simulation–experiment integration
• Monitoring and sentinel agents
• Distributed workflows across institutions
• Scientific systems requiring policy, provenance, and reliability
• Provides the infrastructure layer necessary for real autonomous science: agents that are persistent, routable, robust, and CI-native

​

​

https://academy-agents.org/

Our initial thoughts concerning tech applicability

In brief: LangGraph = controlled reasoning and workflows

AutoGen = conversational agent interactions

Academy = scientific autonomy at scale

​

​

We are developing example code to show how to combine frameworks: e.g., LangGraph for agent logic and Academy for scaling

Agent technologies compared

AmSC and ModCon work to support agent use in GM

AmSC and ModCon work to support agent use in GM

Very much a work in progress! But:

• ModCon BPSW is developing an agent registry to enable discovery and reuse of agent implementations
• ModCon CAF is working to provide templates and examples for agent use in GM-relevant applications
• ModCon CAF is working to scale deployment of LLM instances, LLM agents, and tools on DOE HPC
• ModCon SAFE is working to address safety in agent applications

​

Please tell us of other relevant efforts and missing needs!

​