This page helps Genesis Mission teams evaluate and safely adopt model access, coding assistants, scientific-agent frameworks, and the infrastructure that supports them. Many of the resources described here are designed to be particularly helpful for RFA teams--NNSA-specific onboarding is currently rolling out!
The intended audience includes domain scientists, research software engineers, data scientists, workflow developers, and technical leads. You do not need to build a multi-agent system to benefit from AI.
Start with the lowest level of autonomy that solves the problem. A conventional script, a direct call to an approved model, or a coding assistant is often a better first choice than an autonomous agent.
Before you connect project data or tools
Model access and agent access do not, by themselves, authorize a data use or tool action. Before enabling data access or tools, review the companion pages and confirm the workflow has an owner.
Data Management and Sharing Guidance: Confirm data handling, access, sharing, and retention requirements.
AI Safety and Security: Review security, safety, permissions, and acceptable-use considerations.
Use the Model Access Gateway documentation for onboarding, authentication, project credentials, endpoints, model IDs, and API usage.
Choose your starting point
Use the simplest option that meets the scientific objective. Consider an agentic system when the workflow requires tool use, adaptation, persistent state, long-running execution, or coordination across components.
| What you need to do | Start with | When additional capability is needed |
|---|---|---|
| Ask questions, summarize approved material, extract structure, or test a prompt | A direct Model Access Gateway (MAG) call or approved chat interface | Consider an agentic system when the workflow needs tools, adaptation, or persistent state. |
| Write, explain, refactor, or test code in one repository | A single coding assistant, such as Claude Code, Codex CLI, or OpenCode | Use a scientific agent or workflow framework when the task requires reusable planning, specialist tools, remote resources, or persistent coordination. |
| Capture project conventions and a repeatable procedure | AGENTS.md, CLAUDE.md, and/or a task-specific skill file | Use a structured agentic system when the procedure must manage complex workflows, invoke tools autonomously, or coordinate multiple components. |
| Build hypothesis–plan–execute scientific workflows | URSA | Add specialist agents through MCP, such as MADA, when specialist simulation roles are needed, or use Academy when the deployment requires distributed state and coordination. |
| Set up, manage, and analyze simulation workflows | MADA with MADA Tools | Pair it with a general-purpose scientific workflow agent such as URSA only when the workflow also requires broader hypothesis, planning, or coordination. |
| Build stateful, asynchronous, long-running, or federated agents | Academy | Use Academy deployment patterns with the model, tool, data, compute, identity, and monitoring services required by the workflow. |
| Move data, execute functions remotely, or connect transfer and compute steps | Globus Transfer, Compute, and Flows | Pair Globus with the agent or workflow layer that makes the scientific decisions. |
| Serve a self-hosted model across an HPC allocation | ExaServe | Use it only when the team has a supported site and scheduler, an allocation and approved model weights, model-serving expertise, and an authentication, authorization, rate-limiting, monitoring, governance, and teardown plan. |
| Run a stable, deterministic transformation or simulation pipeline | A conventional script, notebook, workflow engine, or scheduler | Add an agent only for steps that genuinely require interpretation, adaptation, or tool selection. |
Building a practical architecture
A practical architecture combines only the layers required by the scientific objective. The following pages provide starting resources for three classes of tools:
These tools can serve different roles within one system. A coding assistant can help build, edit, and test AI-enabled code. Academy provides middleware for stateful and distributed agents across local, HPC, cloud, or federated resources. It complements, rather than replaces, workflow libraries such as URSA and MADA or underlying frameworks such as LangGraph and Microsoft Agent Framework. Globus can provide access to distributed resources. URSA can orchestrate local scientific workflows and may interface with MADA tools for simulation. ExaServe is a model-serving layer for self-hosted inference on HPC allocations, not an agent framework. URSA and MADA may use MAG for model access and MCP tools for additional capabilities, depending on the architecture.
This is an example, not a requirement. Use only the layers the scientific objective needs.
Resources
Use the following resources when selecting, configuring, evaluating, and operating AI-enabled workflows. Organization-specific links are marked where a Genesis Mission destination is still needed.
Claude Code documentation, Codex CLI documentation, and OpenCode documentation.
Model Context Protocol introduction and the MCP specification.
Globus documentation for transfer, compute, and flows.
NIST AI Risk Management Framework for risk-management concepts and OWASP Top 10 for Large Language Model Applications for common application risks.
Documentation
As with any workflow, AI-enabled workflows need documentation. Documentation helps workflows and their outputs remain reusable, shareable, evaluable, and deployable.
The following templates and resource pages can help you document AI-enabled workflows:
Agents and Skills Guidance: Document repository behavior and repeatable procedures.
Agent card template: Document capabilities, interfaces, authentication, tools, side effects, runtime, memory, intended use, limitations, risk, evaluation, and human oversight.
Model card template: Document underlying model details, data, evaluation, limitations, and governance.
Data Management and Sharing Guidance: Use this when defining approved inputs, access, sharing, and retention.
A reusable agent should not be considered ready merely because it completes a demonstration. It should have an owner, version, documented permissions, evaluation evidence, known limitations, and a defined support and retirement path. Teams may also benefit from a model and agent management plan covering ownership, versioning, evaluation, access, and retirement.
A first successful hour
1. Request access to the Model Access Gateway
Navigate to Genesis Mission Resource Request Page and request access to resources you need for your (RFA) team, including the Model Access Gateway.
2. Confirm access and allowed data
Obtain an American Science Cloud account.
Follow the Model Access Gateway documentation to request access, create a project credential, identify the correct endpoint, and list the model IDs available to your project.
Do not copy an example model ID from a reference card without checking the current model list for your credential.
Confirm which project data may be used before sending the first prompt.
3. Work in a safe project copy
Create a branch, clean clone, container, or disposable workspace. Confirm that you can restore the original state.
4. Select one coding assistant
Teams new to agentic coding should standardize on one tool initially. Choose based on the team's existing model/provider preferences and workflow needs, not on a blanket claim that one assistant is always best.
See Choosing a Coding Assistant for reference cards and additional details.
5. Add project instructions
Use Developing AGENT, Claude, and Skill Files to create a small, testable context file. A useful starter structure is:
# Project purpose
What scientific question or software objective does this repository support?
# Approved inputs
Which files, datasets, endpoints, and data classes may the assistant use?
# Allowed actions
What may it read, create, edit, execute, or call?
# Prohibited actions
What must it never access, change, delete, transmit, publish, or submit?
# Commands
How should it install dependencies, run tests, lint code, and reproduce results?
# Scientific acceptance criteria
What quantitative checks, physical constraints, baselines, or expert reviews define success?
# Artifact locations
Where should generated code, logs, plots, reports, and temporary files go?
Prefer a vendor-neutral AGENTS.md when possible. Add CLAUDE.md only for Claude-specific behavior and a skill file for a narrow, reusable procedure.
6. Run a read-only task first
A safe first prompt is:
Read-only task. Do not modify files, execute commands, access the network, or call external services.
Explain the repository structure, identify the main scientific workflow, list assumptions you are making, and propose three small tasks that can be validated with existing tests or data.
Then allow one small change with explicit files, commands, and acceptance tests. Review the diff and test output before increasing autonomy.
7. Add an agent framework only when the requirement is clear
Write down why a coding assistant, direct model call, or conventional workflow is insufficient. Identify required tools, state, permissions, deployment targets, approval points, failure recovery, and evaluation criteria before selecting URSA, MADA, or Academy.
Troubleshooting and support
Access or model errors: Recheck the current MAG documentation, credential environment variable, endpoint, and model list. Model availability is project-specific.
Unexpected agent behavior: Stop the run, preserve logs and artifacts, inspect the exact tool calls and file changes, restore the workspace if needed, and reduce permissions before retrying.
Scientific uncertainty: Pause automation and involve the domain lead. Do not use model confidence or fluent prose as evidence of correctness.
Security concern: Revoke exposed credentials, stop affected services or jobs when safe to do so, preserve relevant evidence, and follow your institution's security-incident process.
Repository content issue: Open an issue in the RFA launch repository with the page, tool version, operating environment, observed behavior, and non-sensitive error details.
Additional Resources
Several additional resources have been made publically available to help get you started with AI and Agentic workflows.
Explore additional resources here!
Glossary
MAG
Model Access Gateway: a service for requesting and accessing approved models.
MCP
Model Context Protocol: an open protocol for connecting AI applications to tools and data sources.
URSA
A scientific plan-and-execute framework for quickly prototyping local workflows. Use it when a scientist or small team needs explicit hypothesis, planning, and step-by-step execution.
MADA
A framework for specialist agents in simulation workflows. Use it when separate simulation, job-management, retrieval, analysis, or critique roles are needed.
Academy
Agentic middleware for stateful and distributed agents across local, HPC, cloud, or federated resources. Use it for long-running processes, asynchronous communication, remote actions, lifecycle management, or coordination across compute nodes.
ExaServe
A model-serving layer for self-hosted inference on HPC allocations. Use it when a team needs to serve approved model weights on a supported site and has the required operations, security, and governance support.
HPC
High-performance computing: shared systems for large-scale or computationally intensive workloads.
DMP
Data management plan: a plan for how project data will be collected, managed, protected, shared, preserved, and retired.
MMP
Model and agent management plan: a plan covering ownership, access, evaluation, limitations, support, and retirement. Confirm whether this is the preferred Genesis Mission term.
Submit a support ticket for any of the following:
Connect your team with services like data management planning, supercharging your scientific workflows with AI best practices, and cross-cutting AI capabilities.
Have an issue, suggestion, or addition to this content.