Agent Harness

Use GroundX With Your Agent

Once GroundX Agent Harness is installed and your agent is connected to GroundX, you can ask your agent, in plain language, to do real work with your documents. You describe what you want, and the agent runs the GroundX steps for you. You don’t touch any of the machinery yourself.

There are many things you can do this way. Below are three common ones to get you started. They are examples, not the full list.

Before you start, install the harness and connect your agent to GroundX. See Install GroundX Agent Harness and Connect Hosted MCP Tools. Use a GroundX API key that can add and search documents; a read-only key won’t be able to load or search anything. An agent that can’t reach GroundX can explain the steps but can’t run them.

Ask questions about your documents

Hand your agent a document and ask about it, all in plain language:

Set up a GroundX bucket called "my-project", add https://example.com/report.pdf
to it, and tell me when it's ready to search.
What were the Q3 revenue figures in that report, and which page do they come from?

The agent creates a place to store the document (GroundX calls this a bucket), loads it, and waits until it is ready before answering, so it never answers against a half-processed file. Then it searches and answers using what it found. Ask it to “show the exact passages” or “cite the sources” and it points you to the specific text and pages the answer came from. If your agent can read local files, give it a file path instead of a web link and it uploads the file for you using your API key.

Pull structured data out of your documents

When you want a document to come back as structured data your application can use, tell the agent the fields you want (or paste a sample of the JSON you want back) and hand it a few example documents:

Here are three utility statements: <links>. Pull the account number, due date,
total amount due, and service address from each one, as JSON.

The agent designs the extraction, runs it on your documents, checks the result against what is actually on the page, and refines it on its own until the output holds up. Look at what comes back and tell it, in plain language, what is still wrong (for example, “the due date should be the statement date, not the payment date”), and it adjusts. Once it is accurate on your samples, point it at the rest of your documents.

Generate a report across a set of documents

When you have a whole set of documents and the same questions to answer about each one, ask the agent to work through them and assemble a report:

Go through every statement in the "claims" bucket, answer these questions for
each one: [your questions], and put the results in a report with the source for
each answer.

The agent works through the documents in the bucket, answers your questions against each one, and assembles a single report, with every answer linked back to the document and passage it came from. This is how teams build reviews like fraud checks across a stack of statements, or a summary across a folder of contracts.

Prefer to write this in your own code?

If you would rather build any of this into your own application instead of directing an agent, the same steps are available directly:

Next Steps