URSA
Fastest path to a local scientific plan-and-execute prototype
Use URSA when:
One scientist or a small team wants a scientific assistant in a local project workspace.
The workflow benefits from an explicit hypothesizing and planning followed by step-by-step execution.
Built-in scientific agents, retrieval, optional web tools, or general MCP connections match the task.
The team wants to learn what an agentic scientific workflow should do before engineering a distributed service.
Do not start with URSA when:
A deterministic script or workflow engine already solves the task.
The primary requirement is a multi-user production service, durable distributed state, or federated deployment.
The team cannot safely permit file writes or shell execution (URSA will do these things; see below)
Important Operating Guidance:
The execution agent can write files and run shell commands. Use a dedicated or disposable workspace.
A safety agent evaluates each action with settable levels of scrutiny but the user will be responsible for anything done.
Keep web access disabled until the task and data permit network calls.
Do not treat a named agent as a guarantee of production-grade persistence or recovery.
Configure a named URSA group with an allowlist of approved model base URLs when endpoint restriction matters. The default group does not enforce a base-URL restriction.
Add explicit checkpoints, artifact manifests, and validation tests around any consequential workflow.
Additional References:
MADA
Specialist agents for simulation workflows
Use MADA when:
The task requires simulation specialist roles, such as simulation, job management, data retrieval, analysis, or critique.
Existing MADA Tools for schedulers, simulations, workflow management, or other scientific services reduce integration work.
Do not start with MADA when:
The workflow is not simulation-based.
One assistant with one or two tools can complete the task.
The team has not defined specialist boundaries, tool ownership, authentication, ports, timeouts, and failure behavior.
The team expects the planner itself to use configured MCP servers directly. In the current implementation, planner MCP entries are ignored; MCP tools belong on specialist agents, which the planner invokes as tools.
Important operating guidance:
Give each specialist the minimum set of tools and permissions.
Match MADA and MADA Tools server configurations, transports, ports, and authentication settings.
Test each MCP server independently before testing orchestration.
Define how partial failures, long-running jobs, repeated submissions, and conflicting specialist outputs will be handled.
Evaluate both routing quality and scientific task quality; a fluent final answer can hide a failed tool or incorrect delegation.
Additional Resources
Academy
Stateful and distributed agents across federated resources
Academy is agentic middleware, not a peer alternative to libraries like URSA and MADA or the frameworks that underlie them like LangGraph and Microsoft Agent Framework.
Use Academy when:
Agents must preserve explicit state across a long-running process.
Components need asynchronous communication, remotely invocable actions, autonomous control loops, or lifecycle management.
The system must deploy agents across local, HPC, cloud, or federated resources through different executors.
The design requires hierarchical work distribution, remote spawning, monitoring, or coordination across compute nodes.
Do not start with Academy when:
The immediate goal is to try a first prompt, edit a repository, or prototype a short local workflow.
The team does not have software-engineering ownership for deployment, logging, serialization, state recovery, authentication, and operations.
Distribution is a possible future need rather than a present requirement.
Important operating guidance:
Treat every
@actionmethod as a remotely invocable interface and expose only what peers should call.Design control loops to stop cleanly and define lifecycle, retry, timeout, and recovery behavior.
Decide where agent state, messages, logs, and artifacts reside and how long they persist.
Begin from a tested blueprint rather than inventing a federated architecture from scratch.
Pair Academy with the appropriate model client, data service, remote executor, and workflow tools; Academy is the agent runtime and coordination layer, not the model or data system.
Additional Resources
