deepagentsr

deepagentsr is an R-native agent harness inspired by Deep Agents. It provides a high-level create_deep_agent() API with R tools, a virtual filesystem, planning tools, subagents, skills, memory, human approval, permissions, context offloading, and event traces.

The package is designed to use ellmer chat objects for real LLM calls, while the included fake chat model makes tests and examples deterministic.

library(deepagentsr)

search_tool <- deep_tool(
  function(query) paste("mock result for", query),
  name = "internet_search",
  description = "Search a mocked index.",
  side_effects = "read"
)

agent <- create_deep_agent(
  model = fake_chat(list(
    assistant_tool_call("write_todos", list(items = list("Search", "Summarize"))),
    assistant_tool_call("internet_search", list(query = "ellmer R package")),
    assistant_message("ellmer is useful because it provides chat and tool-calling abstractions for R.")
  )),
  tools = list(search_tool),
  backend = memory_backend()
)

result <- agent$invoke("Research ellmer.")
result$text

## [1] "ellmer is useful because it provides chat and tool-calling abstractions for R."

Safety posture

The default backend is in memory. Local filesystem access is an explicit capability grant through filesystem_backend(root_dir), which maps virtual paths into a configured root and blocks traversal and common secret-like paths. Shell execution is not included in the default runtime.

Optional integrations

ellmer is the intended model and tool-calling substrate. MCP, RAG, Shiny chat, and aisdk interop are optional helpers so the core package stays small while extension points remain available. The guarded live test path exercises OpenAI tool-calling across GPT 4.1 and GPT 5 family models when credentials and model access are available.