How to Create a Chat Agent in Eve

Overview

A chat agent is a saved prompt, or a saved sequence of prompts, that runs in one click and returns a conversational response: case analysis, summaries, checklists, recommendations. It is built once and available on any matter, to you or to everyone at the firm.

Agents are created from the Agents tab in the left navigation.

Before You Start

• Access to the agents feature.

• A clear idea of the task, ideally one already run successfully as a prompt.

Steps

1. Test the prompt in a matter first. Run it in a normal conversation and confirm it produces what you want. Wording is easier to refine in chat than in the agent editor, and whatever you land on becomes the instructions.

2. Open agent creation. Click Agents in the left sidebar, then Create Agent in the upper right. You can also build from a prompt you have already run: hover over that message in the chat and click the magic wand icon beneath it. Agent creation opens with the prompt loaded.

3. Select Chat Agent as the output type.

4. Write the first instruction. Write it the way you would write a prompt in a normal conversation. To make it more specific, try to ask for at least these four things:

   • What to produce. Format, length and structure. "Return a table with columns for date, provider, treatment and cost" produces more consistent results than "summarize the medical records".

   • What to include. The case details, legal standards or data points that must appear every time.

   • What to leave out. Anything that should never appear, stated explicitly. Exclusions are easy to forget and they prevent the most annoying failures.

   • How to use attached reference files. If Central Library documents will be attached, say what to do with them, such as "use the objection bank to draft applicable objections for each request".

5. Add sequential steps with + New Message. When the workflow has several stages, add another prompt rather than packing everything into one. Each message runs in order and builds on what came before. Your prompts appear in sequence in the creation window.

Splitting a workflow into steps does two things: it gives Eve room to work through each stage properly instead of racing through one dense instruction, and it exposes the reasoning at each point, so when output is wrong you can see which step went sideways instead of rewriting everything.

A discovery analysis works better as three messages than one: identify deficient interrogatory responses, then identify documents missing from production, then format both into a meet-and-confer letter.

Click Next when the sequence is complete.

6. Add additional agent capabilities. See 📄 New User Training Video #7: Chat Agents and Advanced Agent Functionality.

7. Review the name, description, tag and visibility. Eve generates a name and description from the instructions. Read them before saving, since they are what colleagues see when deciding whether to use the agent.

• Change the name if it describes the matter rather than the job. "Intake Call Summary" beats "Johnson Intake v2".

• Change the description if it does not say which documents to select before running. That omission causes the most common failure, someone running the agent without the right context selected, and auto-generated descriptions do not always catch it.

• Set the tag. An agent carries one tag, so pick the one that fits your firm's convention. See 📄 How to Organize Agents with Tags in Eve.

• Set the visibility. Only me keeps it private; Shared makes it available to everyone at the firm.

8. Save the agent. It appears in the Agents tab and is available in any matter.

9. Run it on a matter. Open the matter, select the context documents the agent needs, open the chat panel and select the agent. See 📄 How to Run a Chat Agent in Eve.

Refine After the First Real Run

The first version is rarely the final one. Run it on a matter where you know what good output looks like, compare, and edit where it fell short. Because the instructions are broken into steps, you can usually fix the one message that is underperforming instead of rewriting everything.

Asking Eve directly is worth trying: "how could I ask this task of you in the future?" or "how would you rewrite these instructions to get a better result?" often returns a more precise version of your own instructions. Expect two or three rounds. See 📄 How to Edit Agents in Eve.

Running an Agent Automatically

Chat agents run manually by default. A trigger makes one run on its own, on matter update or on a monthly schedule, with nobody pressing a button. Writing and testing trigger conditions is covered separately in 📄 How to Configure Trigger Conditions for Auto Agents.