Chat Agents
Last updated: September 22, 2026
Overview
A chat agent is a saved prompt that can be run on any matter. It holds the instructions for a task somebody at the firm does repeatedly, so the work starts from a click instead of a blank message.
Chat agents live in the Agents section of Eve and run inside a matter, against the documents in that matter. Every workspace comes with a set of example agents already built, and firms add their own.
The name describes the task. A chat agent does what a person would do in Talk to Eve, following instructions that were written and tested once instead of retyped every time.
Where Chat Agents Live
Open Agents from the left navigation, then See all agents for the full list.
Two things determine what appears there:
Visibility settings. Each agent is either private(visible only to the person who created it), or shared with the firm. Shared agents are how a workflow one attorney figured out becomes the way the whole firm does that task.
Type. Chat agents sit alongside drafting agents, auto agents, and communication agents in the same list. You can use the Agent Type filter to look at just Chat agents, if necessary.
Running a Chat Agent
Open the matter to run it against.
In the center of the matter, above your four recommended agents, select See all agents and click the agent you wish to run.
Set the matter context. Everything left checked is what Eve reads for this run. Everything unchecked may as well not exist for the duration of the task. If a folder has nothing to do with the question, uncheck it; if it is unclear, leave it in.
(Optional) Add Central Library or Research context if the task should account for firm reference material such as a style guide, legal research, or a standing set of instructions.
Click Run agent.
The response arrives in the conversation the same way any other Eve response does, with citations back to the source documents. The conversation stays open afterward, so a follow-up question, a narrowing instruction, or a correction goes in the same thread rather than in a new one.
Single-Step and Multi-Step Agents
A chat agent can hold one instruction or several. Several instructions are called instruction steps, and Eve completes each one and returns a response before starting the next.
For example, a witness analysis agent with three instruction steps runs like this:
Compile every potential witness in the matter.
Identify the five best candidates for deposition.
Draft an outreach email to each of the five.
Each response appears before the next step begins, so the reasoning is visible in order rather than arriving as one block at the end. It also means each step can be read and checked on its own, which matters most on the steps that feed the ones after them. If the witness list in step one is wrong, nothing downstream of it is worth reviewing.
Instruction steps are also the answer to a workflow that spans what would otherwise be two agents. One agent cannot be set to run off another agent's output, so a task that needs both halves belongs in a single agent with steps.
When a Chat Agent Is the Right Tool
The signal is repetition, not importance. Having typed roughly the same series of prompts on three different matters is enough. That workflow is a chat agent, and the next person who needs it should not have to reinvent the prompts.
Some things are better left as chat:
A one-time question. Settling a specific point on a specific file does not need to be saved.
A small variation on an agent that already exists. Run the existing agent, then ask for the variation as a follow-up in the same conversation.
A task whose instructions change every time. If the prompt has to be rewritten for each matter anyway, saving it adds a step instead of removing one.
What a Chat Agent Can't Do
It does not run on its own. A person opens a matter and runs it. Work that should start on a schedule or in response to something happening in a case is an auto agent instead. See 📄 Auto Agents and Triggers.
It does not contact anyone outside the firm. Anything that sends to a client, a provider, or opposing counsel is a communication agent, which carries its own approval step. See 📄 Communications Agents Overview
It only knows what is in context. A document that was unchecked in matter context, or never uploaded to the matter, cannot be read or cited.
Its output is a first draft. Review it, and use the citations to check the claims that matter before anything leaves the firm.
It cannot be chained to another agent. Use instruction steps in one agent.
How Chat Agents Compare to the Other Agent Types
You do | Eve does | What it is called |
|---|---|---|
Ask, inside a matter | Runs the saved instructions and returns a response | Chat agent |
Ask, inside a matter | Produces a document for editing | Drafting agent |
Set up a condition once | Runs when the condition is met, ready for review after | Auto agent |
Approve | Sends a message to someone outside the firm | Communication agent |
Key Terms
Chat agent: A saved prompt, run manually inside a matter, that can be reused on any case.
Instruction steps: The individual prompts inside a multi-step chat agent. Eve returns a response for each one before moving to the next.
Agents: The section of Eve where agents are stored, organized, shared, and run.
Private agent: An agent visible only to the person who created it.
Firm-shared agent: An agent available to everyone in the firm's workspace.
Matter context: The documents and files from a matter that Eve reads during a run. Unchecked files are excluded entirely.
Central Library context: Optional firm reference material Eve accounts for during a run, selected in addition to matter context.