Agent Prompts

Agent Prompts is the DataMind staff panel that keeps the register of agents: the named identities whose system prompts the platform adds to the requests that name them. Use it to register an agent, keep its display name current, stop or resume it, and open its prompt history.

The two views

The panel has two views, Agents and Observability. It opens on Agents; both views are deep-linkable, so a link can open the register or the audit directly, and the address also records the agent you have open.

• Agents is the register. Its header shows Agent Prompts and the number of agents; each row shows the display name, the key, when the agent was last updated and an Enabled or Disabled badge. Selecting a row opens the agent's page. When the register is empty it reads No agents yet. If the register cannot be read at all, the area shows the error message instead.
• Observability is the injection audit: the requests whose system prompt did not come from a published version. It is not scoped to one agent, and it has its own page, Injection audit.

The register's toolbar holds Add agent and Publish from local files. The second button appends prompt versions from the bundled prompt files; Agent prompt versions describes what it compares and what it writes.

Register an agent

1. Click Add agent in the toolbar of the register.
2. Enter the Key: the value callers send as the agent field, in kebab-case, for example agentic-mind. It cannot be changed after creation.
3. Enter the Display name: the label shown here and in the agent lists, for example DataMind OS Assistant.
4. Click Add agent in the dialog. A successful save shows Agent created.

The dialog states what comes next: "Register a new agent identity. Publish a version afterwards to give it a prompt." A new agent starts Enabled, and until its first version is published the platform serves its built-in prompt. Both fields take up to 100 characters.

The form checks itself before saving: a missing name reads Display name is required, and a key that breaks the kebab-case rule reads Key must be kebab-case: lowercase letters/digits separated by single hyphens. If the save itself fails, Create failed shows the reason.

Edit an agent

1. Open the agent from the register, then click Edit.
2. Change the Display name. The Key is read-only, and the dialog says why: "The key clients send cannot change."
3. Click Save. A successful save shows Agent updated.

Closing the form with unsaved edits asks Discard unsaved changes? and warns "Your edits will be lost."; Discard closes the form without saving.

Disable or enable an agent

The switch in the agent's header, Agent enabled, controls whether the agent serves. There is no delete on this panel.

• To stop an agent serving, switch it off. A confirmation asks Disable <displayName>? and explains: "Disabling removes the published prompt from the proxy - requests for this agent fall back to the built-in default prompt. The version history is kept." Confirm with Disable.
• To serve its published prompt again, switch the agent back on. Nothing is republished; the timeline stays as it was.
• A disabled agent keeps its row in the register with a Disabled badge, and its page keeps the full timeline.

Open an agent's prompt history

Selecting an agent opens its page: the display name and key in the header with the switch, then Edit and Publish new version, and the version timeline below, newest first.

• The timeline header reads "Version timeline - newest first. Versions are immutable; publishing appends."
• Above it, the page notes which deployments receive which version: each client deployment receives the newest version dated on or before its build date, and older deployments keep their era's prompt.
• Until a first version exists, the timeline reads "No versions published yet - the proxy serves this agent's built-in fallback prompt until v1 is published.", with a Publish first version button.
• The back button is Agents; if the agent cannot be loaded, the page shows the reason with Back to agents instead of the timeline.

Publishing a version, scheduling it and editing its weights are covered on Agent prompt versions.

Questions, answered

Can I change an agent's key after it is created?

No. The key is fixed at creation, because callers send it as the agent field of their requests; only the Display name can change. If you need a different key, register a new agent and disable the old one: its history stays on its page.

What happens if I choose a key that is already in use?

The create is refused: Create failed shows the reason, naming the clashing key, for example Agent "agentic-mind" already exists. Keys are unique, so pick another one.

Can I delete an agent?

No. Disabling is the lifecycle: switch the agent off to stop it serving, and its row and version history stay.

What happens to requests while an agent is disabled?

They are served the built-in default prompt instead of the agent's published prompt, as the confirmation states. Switch the agent back on to serve its published prompt again.

Does disabling an agent keep its version history?

Yes. Disabling changes only whether the agent serves; the versions are kept, and the row stays in the register with a Disabled badge. You can switch it back on at any time.

Which prompt does a new agent serve before its first version is published?

The built-in fallback prompt: the agent's page states this until v1 exists. Publish first version gives the agent a prompt of its own, and the full publishing flow is on Agent prompt versions.

What is the difference between the Agents and Observability views?

Agents is the register and each agent's prompt history. Observability is the injection audit, which lists the requests whose system prompt did not come from a published version; it is not scoped to one agent. See Injection audit.

Can a company role open Agent Prompts?

No. It is a DataMind staff surface: no company role or permission opens it.