AgentMaker

AgentMaker is the copilot in the agent editor. Describe the result you want in plain language, and it analyzes the available configuration and returns a proposal for review. A prompt change is not saved merely because it appeared in the conversation.

This page covers AgentMaker for agents. The automation editor uses another mode of the same copilot to build a flow; see Building the automation with AgentMaker.

Before you start

You need permission to edit agents (agents.write) and must use the editor on a desktop-sized screen. The panel is not displayed at smaller widths.

Save any pending manual changes before starting a conversation. The context sent to AgentMaker uses the last saved prompt, not text that is still unsaved in the editor. This precaution prevents a proposal calculated from the previous version from replacing your draft.

Prompts can contain up to 32,000 characters in the editor, but AgentMaker receives only the first 8,000 characters of the saved prompt. For long prompts, request small changes and inspect the entire text before applying a full rewrite.

What it can analyze

In agent mode, AgentMaker receives:

• the saved prompt and diagnostics about its structure;
• tools active on the agent and tools available to the organization;
• the names and descriptions of linked knowledge bases — not their item content;
• relevant agent settings and capabilities of the selected model;
• the current administration navigation map, so it can tell you where to change a setting.

With this context, it can stage a section rewrite, add a section, replace the entire prompt, suggest tools, or explain a configuration problem. It does not test the agent or guarantee that a proposal will produce the expected behavior.

How to open it

The editor has four entry points:

1. After you create an agent manually, the editor opens AgentMaker automatically on desktop.
2. Select AgentMaker in the editor header. This button works from any section.
3. In the Prompt section, select Edit Prompt. This opens the text editor and the panel together.
4. Under Insights, select Ask AgentMaker. The panel opens and the suggested question is sent automatically.

The X closes only the panel. In the Prompt section, selecting Preview also closes it and returns the text to read mode. History opens previous prompt versions and is a separate flow.

The Test Agent panel and AgentMaker cannot stay open at the same time: opening one closes the other.

Review and apply a proposal

1. Enter a specific objective, such as: “make the tone formal and prohibit discounts without approval.”
2. Wait for the response. You can interrupt generation with the stop control.
3. When changes can be applied, inspect the Proposed improvements card and expand the sections to compare the text.
4. Select Apply all or Discard. An explanatory response without staged changes does not show these buttons.
5. Review the editor and select Save to persist the new prompt. Use Discard in the editor bar if you do not want to save the draft.
6. Test the agent in a controlled conversation before publishing or using the new configuration.

Applying a proposal replaces the prompt draft in the form; it does not merge with parallel manual edits. For large changes, apply and save one proposal at a time.

Tool changes

AgentMaker can propose enabling or disabling native and Composio tools. When you select Apply all, those tool changes are persisted immediately; the prompt portion remains a draft until you select Save.

Review the Tools section after applying. Some native tools still require their own configuration — for example, selecting which bases a search can query. MCP servers and tools are part of AgentMaker’s context, but it cannot enable or disable them from the review card.

If applying a batch fails, do not assume a complete rollback: tools processed before the error may remain changed. Check Tools, correct the configuration, and only then decide whether to try again.

For settings it cannot change directly, AgentMaker displays a recommendation and the relevant interface path. You remain responsible for opening the indicated section and confirming the change.

Conversation history

After a response finishes, the browser stores the conversation by user, agent, and AgentMaker mode. Reloading or returning to the same agent in the same browser normally restores the messages and a pending proposal.

Reset chat asks for confirmation, clears the history displayed in that browser, and starts a new conversation. It does not change the saved prompt, undo tools that were already applied, or constitute a request to delete conversation records retained by the service.

The help button, labeled What AgentMaker does, summarizes the capabilities of the current mode and links to this documentation.

Best practices

• save the manual draft before requesting an analysis;
• specify the role, audience, tone, rule, and outcome you want to change;
• prefer one verifiable improvement at a time;
• read the full diff, especially when the prompt exceeds 8,000 characters;
• check tools and recommended settings before saving the prompt;
• test normal cases, limits, and refusals after every change.