How to Tell Auto Agents Which Documents to Focus On

Overview

An auto agent is a chat or drafting agent with a trigger on it, so it runs on its own across matters with nobody there to steer it. That changes what the instructions have to do. When an agent is run by hand, a person picks the matter, selects the context, and knows what the task is about. When it runs on a trigger, the instructions carry all of that on their own.

Instructions that do not name the documents leave Eve to review everything in the matter, and the output comes back broad and shaped by files that may have nothing to do with the task.

For setting the trigger itself, see 📄 How to Configure Trigger Conditions for Auto Agents.

How an Auto Agent Decides What to Read

Three things determine what a run draws on:

• The instructions. The primary signal, and the only one under direct control.

• The run condition. For example, a run started by new medical records carries the implication that medical records are what matter.

• What is actually in the matter. No instruction produces a good result against documents nobody uploaded.

Before You Start

• Firm admin access, which is required to create or edit auto agents.

• The agent created, or open in the builder.

• A clear answer to what the agent is for, in terms of document types, folders, or time windows.

Writing the Instruction

• Name the document type. Use the titles, folders, or data points your matters actually use. "Focus on medical records," "analyze bills and invoices," "review adjuster correspondence," "summarize court filings." An instruction to review relevant documents tells Eve nothing, because relevance is the thing being decided.

• Say whether recency matters. Without a time window, everything in the matter is included the same, including records from two years ago. Write "documents uploaded in the last 7 days," "the most recent medical records," or "correspondence received since the last run" when the point of the agent is what just arrived.

• Use exact folder names. If matters follow a consistent folder structure, name the folder: "documents in the Medical Records folder," "files in Discovery Received." Folder names must match exactly. DISCOVERY will not match Discovery Received. Check the naming convention across several matters before writing the instruction, because an agent running firm-wide meets every variation your firm has ever created.

• Say what to ignore. Exclusions do as much work as inclusions, especially on an agent that runs often. "Do not include correspondence between the client and this firm," "exclude internal case notes," "ignore billing records." Left unsaid, that material gets read and shows up in the output.

Pair the Condition With the Instruction

The run condition narrows which matters the agent looks at. The instruction narrows which documents it reads inside them. They work best when both describe the same intent.

Set the Run condition on the agent to scope the matters, then write instructions that pick up where the condition left off:

A condition that fires on medical records paired with an instruction to summarize the matter wastes the specificity the condition gave you.

A Complete Instruction

An example of what instructions may look like for an agent that runs when new records arrive:

Review medical records uploaded to the Medical Records folder in the last 7 days. Summarize the treatment documented in them and flag any gap of 30 days or more between visits. Do not include internal case notes, billing records, or correspondence between the client and this firm.

Document type, recency, folder, exclusions, and a specific task. Compare that to "review new medical records and summarize," which leaves Eve to decide what counts as new, where to look, and what the summary is for.

Test Before Turning It On

1. Run the agent manually on two or three matters that represent the range of what it will meet, including one messy file rather than three clean ones.

2. Read the output against the matter and check which documents it actually drew on.

3. Refine the instruction wherever the output pulled in something irrelevant or missed something it should have found. Missed documents usually mean the folder name or document type is wrong. Irrelevant content usually means an exclusion is missing.

4. Repeat until two consecutive runs land where they should, then enable the trigger.

What Instructions Cannot Fix

• A folder that does not exist in every matter. An instruction pointing at Medical Records does nothing on a matter where that folder is named something else, and the run will quietly produce less than expected rather than reporting a problem.

• Documents nobody uploaded. An auto agent can only read what is in the matter at the moment it runs.

• An instruction doing two jobs. An agent asked to summarize records, draft correspondence, and flag deadlines in one instruction does all three at medium quality. Split it.

• Review. Output from an auto agent lands for a person to read, the same as anything else Eve produces, and the citations are there to check it against.