AutoPrompt

What it is

Experimental native tool that lets the agent edit its own prompt during conversation. With AutoPrompt active, the agent can list, read, rewrite, add sections, or replace its entire prompt — and changes take effect on the next message.

It’s sandbox mode: no review, no approval, no audit trail. Changes are applied directly to the agents.prompt column in the database.

When to use

You’re prototyping an agent and want to tune the prompt by chatting (“be more formal”, “now always cite the source”, “remove the X instruction”).
Quick prompt engineering experiments without opening the editor manually.
Demoing how the agent behaves with different prompts.
Meta agents that calibrate from user feedback in a supervised session.

Do not use in production with agents customers can reach or in environments where the prompt is part of compliance.

Configuration

In the agent’s Tools tab, click + Add Tool.
Under NATIVE TOOLS, pick AutoPrompt.

No parameters, no configuration. The tool is active immediately.

How the agent invokes

The agent decides to call this tool when the user explicitly asks:

“adjust your prompt to X”;
“remove the rule about Y”;
“add a section about Z”;
“show me your current prompt” (reads without modifying).

The agent confirms the change in its reply and, from the next turn, runs with the new prompt.

Limitations and notes — read before enabling

No audit. No log of what changed, by whom, when. The original prompt is lost on every rewrite.
No review. Anyone chatting with the agent can trigger the edit — just by asking. In public conversations, this is a real risk.
Changes apply to all conversations. The edit changes agents.prompt directly. Parallel conversations on the same agent use the new prompt on the next turn.
Declared sandbox. The UI makes it clear this tool has no safety guarantees — use it on test agents, not in production with the public.
For safe edits, prefer editing the prompt manually in Agents → [agent] → Prompt, where you see the diff before saving.