Skip to content

Agents

URSA agents are reusable behaviors that can chat, plan, execute, search, maintain persistent research artifacts, reason over documents, refine prompts, or interact with external tools.

The interactive command-line interface exposes common agents by short names, including:

  • chat
  • plan
  • execute
  • prompt
  • arxiv
  • web
  • deep_review
  • hypothesize

Additional agents may be available depending on optional dependencies and configuration, such as DSI, LAMMPS, and recall/RAG-related workflows.

Common agents

  • Chat Agent: general conversational agent with local workspace/file/experience tools and optional web/literature tools.
  • Planning Agent: decomposes requests into a structured plan and reflects on it before returning the final plan.
  • Execution Agent: carries out workspace-changing tasks with code, file, command, experience, optional search/RAG/MCP tools, and a structured review-until-complete loop before recap.
  • Prompting Agent: helps refine prompts and task descriptions for downstream agents.
  • Acquisition Agents: shared acquire-then-summarize/RAG agents for external sources:
  • ArXiv Agent: searches arXiv, downloads PDFs, extracts text, and summarizes or indexes papers.
  • OSTI Agent: searches OSTI records, resolves available full text or landing-page content, and summarizes or indexes records.
  • Web Search Agent: searches the open web, retrieves HTML/PDF content, and summarizes or indexes acquired pages.
  • Deep Review Agent: runs an iterative adversarial review with solution, critique, and competitor/stakeholder phases, then synthesizes a final answer/report.
  • Hypothesizer Agent: maintains a persistent Markdown hypothesis-space artifact that other agents can read from the experiences store.

Web and external information

Web/search tools are opt-in for information-control reasons. For CLI sessions, enable them with:

ursa --use-web

With --use-web, tool-capable agents that support web access, such as ChatAgent, ExecutionAgent, DeepReviewAgent, and PromptingAgent, receive web/search tools according to their implementation. Without it, they should rely on local workspace, experience, RAG, MCP, or other explicitly configured tools.

Composing agents

For work that benefits from multiple roles, URSA environments compose agents behind one invoke(...) interface. Use an Agent Team when one PI should delegate to specialists and synthesize a coherent answer. Use an Agent Symposium when several agents or nested teams should work independently, review one another, revise, and then produce a final synthesis.

Safety note

Agents with tools can access local files, run commands, or make network requests depending on configuration. Use dedicated workspaces for execution-heavy workflows and review generated commands/code when appropriate. See Sandboxing and information control.

Agent documentation

Use the individual agent pages for current class-level details and examples.