> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agents.kolena.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Leonard

> Work with Kolena's built-in AI assistant to build, explain, and improve your Agents

Leonard is Kolena's AI assistant, built into the platform. He works alongside you in a side panel:
he can read the Agents in your Workspace, explain what a Prompt did, propose edits for you to review,
build new Agents, and answer how-to questions with links to this documentation.

<Note> This feature is in beta and is enabled per Organization. [Contact Kolena](mailto:support@kolena.com) if you'd like it turned on for your Organization.</Note>

## Opening Leonard

Click **Work with Leonard** in the top navigation, or press `Ctrl+Shift+L` (`⌘⇧L` on Mac).
The panel opens beside your work, so the Agent, Run, or Prompt you were looking at stays visible.

Leonard knows which page you have open, so you can refer to what's in front of you (for example: "why did this
Prompt return null?", "add a Prompt here for the lease end date", "how many credits did this Run
cost?") without pasting ids or instructions into the chat.

## Examples

<div style={{ display: 'grid', gridTemplateColumns: '1fr', gap: 24 }}>
  <div>
    <video src="https://mintcdn.com/kolena/TmaWUI5ZFG8WLSC3/assets/leonard-build-agent.mp4?fit=max&auto=format&n=TmaWUI5ZFG8WLSC3&q=85&s=f1905ee2a10613ea9e5a3a8977785671" poster="/assets/leonard-build-agent-poster.webp" controls muted loop playsInline preload="metadata" controlsList="nofullscreen" disablePictureInPicture style={{ width: '100%', borderRadius: 8 }} data-path="assets/leonard-build-agent.mp4" />

    <p style={{ marginTop: 6, fontSize: 14 }}><strong>Building an Agent.</strong> Attach the workbook your team fills out by hand, and Leonard creates the Agent and proposes the Prompts that produce it.</p>
  </div>

  <div>
    <video src="https://mintcdn.com/kolena/TmaWUI5ZFG8WLSC3/assets/leonard-edit-prompt.mp4?fit=max&auto=format&n=TmaWUI5ZFG8WLSC3&q=85&s=9d931e6c89c8f404702c25e5f51dcd77" poster="/assets/leonard-edit-prompt-poster.webp" controls muted loop playsInline preload="metadata" controlsList="nofullscreen" disablePictureInPicture style={{ width: '100%', borderRadius: 8 }} data-path="assets/leonard-edit-prompt.mp4" />

    <p style={{ marginTop: 6, fontSize: 14 }}><strong>Fixing a Prompt.</strong> Ask why a value came back wrong; Leonard traces it to the instruction and proposes the edit for you to accept.</p>
  </div>
</div>

## What Leonard can do

* **Build Agents:** Describe the outcome; Leonard creates the Agent and proposes its Prompts.
* **Explain and improve Prompts:** Review an instruction, add a missing field, or fix a Prompt returning the wrong value. Edits are surgical. He can also edit **Agent Guidelines**.
* **Connect Prompts and narrow documents:** Chain Prompts so one reads another's output, and set Document Filters so a Prompt only sees the documents that matter.
* **Explain results:** He reads a Run's results and its source documents. Verify an extracted value, compare two files, or find something no Prompt captures.
* **Search your Workspace:** Ask which Agents exist, where one lives, or which Runs it has.
* **Answer how-to questions:** He searches this documentation and links the page he used.
* **Report credit usage:** Workspace, Organization, per-Agent, or per-Run (limited to usage you can view).
* **Create files to download:** `.xlsx`, `.docx`, `.pptx`, or `.pdf`, attached to the conversation.
* **Start a Validation:** Against an Agent's [Ground Truth](/how-to/measuring-accuracy), after you confirm.

## Nothing changes until you accept

Leonard never edits an existing Agent directly. Every change arrives as a **Proposed Changes** card
in the conversation, listing each Prompt being added, modified, or removed.

* Click any item to open the full diff and see the instruction and settings before and after.
* **Generate a preview** to run the proposed Prompt against a real Run and see the value it produces,
  side by side with the current value, before you commit to it.
* **Accept all** applies the update; **Reject all** discards it. If part of an update isn't right,
  tell Leonard what to change and he'll propose a revised version.

Leonard can create a **new, empty** Agent when you ask him to build one but all
Prompts still arrive as a proposal for you to review.

<Note>
  Accepting a Prompt change marks existing results for that Prompt out of date. Regenerate them from
  the Agent when you're ready — Leonard cannot create Runs or re-run existing ones.
</Note>

## Attachments and dictation

Attach files to a message when a screenshot or sample document is the fastest way to explain what you
want: up to 10 files per message, 10 MB per file and 20 MB in total. Leonard keeps them for the whole
conversation, so you can refer back to a file you attached earlier without uploading it again.

You can also dictate a message with the microphone in the chat input instead of typing it.

## Chats and history

Each conversation is saved. Use the history menu to search and reopen past chats, or delete ones you
no longer need; **New chat** starts fresh. Long conversations are summarized automatically as they
grow, so Leonard keeps track of what you're working on. Very long chats eventually reach a limit and
ask you to start a new one — start a new chat per task for the best results.

A thumbs up or down on any response (with an optional comment) tells us what to improve.

## What Leonard can't do

* Change an existing Agent without your approval.
* Create Runs, upload documents to a Run, or re-run an existing Run.
* See anything you can't see. Leonard runs with your permissions in your current Workspace.

## Permissions and privacy

Leonard enforces exactly the same access rules as the rest of the platform. He can only read Agents
you have access to, can only propose Prompt changes on Agents you may edit, and can only start a
Validation if your role allows it — including under
[fine-grained access control](/fine-grained-access-control).

Conversations with Leonard are **not billed** and do not consume your credits. As with the rest of
the platform, your data and your interactions are never used to train AI models — see
[Data Privacy & Security](/faq/security).

## Tips for better results

* **Work from the page you're on.** Open the Agent, Run, or Prompt first, then ask — Leonard resolves
  "this Prompt" or "this Run" from what you're viewing.
* **Say what's wrong, not just what to write.** "This returns the tenant's name instead of the
  landlord's" gets a better edit than "rewrite this Prompt".
* **Point at a Run when accuracy is the problem.** Leonard can read the source documents and see what
  the Prompt actually had to work with.
* **Batch related edits.** Ask for all the changes you want in one message and review them as a
  single update.
* **Preview before accepting** anything you're unsure about.
* **Start a new chat** when you switch tasks.