Document Research Agent

This is an example recipe, not a core Little Harness concept. Use it as a starting point for agents that work across user files, company docs, contracts, PDFs, notes, or knowledge bases.

Document research is a good first Little Harness use case because it benefits from tools, files, artifacts, and optional user memory.

Shape

const harness = createHarness({
  host: localHost(),
  model,
  system: "Help the user search, compare, and summarize their documents.",
  tools: { searchDocs, readDoc, summarizePdf },
  skills: [skill("./skills/domain-glossary")],
});

Tool Set

Start with three tools:

• search_docs finds candidate documents.
• read_doc fetches a selected document or page range.
• summarize_pdf creates a structured artifact for long PDFs.

Large extracted text should become artifacts:

{
  preview: "Contract PDF extracted. Sections: term, renewal, termination...",
  artifactPath: "/artifacts/docs/acme-contract/extracted.md",
  bytes: 239012
}

Optional User Memory

Add a Persistent Dir when a user wants durable notes across sessions.

persistentDirs: [
  {
    harnessDir: "/persistent/user-notes",
    load: ({ extraBody }) => notes.loadFolder(extraBody.userId),
    store: ({ extraBody, changes }) =>
      notes.applyChanges(extraBody.userId, changes),
  },
]

The agent can then write resolved terminology, preferences, and long-running research notes without mixing them into every prompt.

Export Boundary

Generated summaries should stay as artifacts until the user asks to save or share them. Use explicit tools such as save_summary, create_doc_comment, or send_draft.