AutoPrompt

What it is

AutoPrompt is the native tool an agent uses to read and improve its own system prompt from what it learns in real conversations. It is meant for correcting wrong or outdated guidance and for adding a rule that will remain useful in future conversations.

The agent can list and read the current sections, rewrite a section, or add a new one. The configuration decides whether a change becomes a proposal for review or is applied automatically.

AutoPrompt is not conversation memory or a knowledge base. A request that only applies to the current reply should be handled without changing the prompt. Facts, catalogues, and reference content belong in knowledge bases or context blocks.

Add and configure

You need permission to edit the agent (agents.write). On an already saved agent:

1. Open Agents, select the agent, and go to Tools.
2. Select Add Tool.
3. Under Native Tools, choose AutoPrompt.
4. Choose the mode and save.

A new AutoPrompt starts in It proposes, someone on the team approves. An older installation that does not have a mode field preserves its previous, automatic behaviour. The card summary shows the effective mode; open the configuration and check it before exposing the agent to customers.

Mode	What happens	When to use it
It proposes, someone on the team approves	Rewriting or adding a section creates a proposal. The prompt does not change until approval.	Recommended default, especially for external agents or sensitive rules.
It changes right away, no approval	The change is written directly and takes effect on the next turn.	Only for an internal agent, with controlled use and accepted risk.

There is no “off” mode. To prevent further reads and changes by the agent, remove AutoPrompt from the Tools section.

When the agent should use it

The tool contract itself tells the model to act when a conversation shows that its instructions are wrong, outdated, or incomplete, for example:

• someone corrects a durable behaviour rule;
• the agent notices it answered from outdated guidance;
• guidance is missing for a recurring situation.

An explicit request such as “change your instructions” does not guarantee the change will be accepted: the model still decides whether the information is durable and belongs in the prompt. Do not use AutoPrompt for a one-off request or to store facts that should be searchable in a knowledge base.

Review a proposal

When at least one proposal exists, the agent’s Prompt section shows the Improvement suggestions button and the pending count. Open it to:

• compare the previous and proposed text by section;
• read the reason and open the conversations that originated the proposal;
• Approve, Edit and approve, or Reject;
• switch between Waiting for you and Already reviewed.

The reviewer needs the same agents.write permission required to edit the agent. Approval applies the section and records the change in the prompt’s History; rejection prevents the same proposal from returning for 30 days.

Equivalent proposals are grouped and show how many people requested the change. If the section is edited after the proposal, approval does not overwrite the newer text: the proposal becomes Out of date, and the agent can propose it again from the current version.

In reviewed mode, AutoPrompt works one section at a time. Whole-prompt replacement is refused because it would not produce a readable diff. Structure the prompt with Markdown headings to enable small reviews. If the document has no sections, edit it manually or reorganize it before using this mode.

Automatic mode

In It changes right away, no approval mode, the agent can also replace the whole prompt when a section-level edit is not suitable. Every change guides the agent’s conversations from the next turn onward, including parallel conversations.

History keeps up to the 20 most recent versions and lets you compare or restore one. Automatic changes also create a version, but they can currently be shown as an initial version instead of identifying AutoPrompt as the source. History therefore does not replace prior review or conversation monitoring.

Automatic mode accepts a change chosen by the model inside a conversation, without a second human authorization. A persuasive or malicious message can try to turn a temporary instruction into a permanent rule. Use this mode only when everyone who can talk to the agent is trusted and an unwanted change can be detected and reverted quickly.

Organization context

It can also suggest changes to the blocks it uses is independent from the prompt mode and starts disabled. When enabled, the agent can list the context blocks attached to it and propose rewriting or adding a section.

This scope has three important limits:

• blocks that are not attached to the agent are neither listed nor accepted;
• a block change always requires review, even when the agent’s own prompt is in automatic mode;
• approval affects every agent that uses that block.

Block proposals are reviewed under Organization context by someone with administrative permission. Check the diff and the source conversation before approval because the text is shared.

Safe-use checklist

• Prefer reviewed mode for any customer-facing agent.
• Write the prompt as small Markdown sections with unique headings.
• Before approval, check that the rule is durable, belongs in the prompt, and does not contradict another section.
• Use Edit and approve when the intent is right but the wording still needs adjustment.
• Review History after unexpected behaviour and restore an earlier version if needed.
• Remove AutoPrompt when the agent no longer needs to improve its own instructions.