Agents

An agent is a model with instructions, the collections it grounds on and, optionally, the tools it may call, the skills it may read and the other agents it may hand work to. Build › Agents lists the tenant’s agents; each one has its own page. Builders create and change agents here; an auditor reads them.

An agent is internal to the tenant. It has no URL and no key of its own: publishing it as an assistant is what makes it reachable by people, and nothing an agent does leaves the platform except the tool calls you grant it.

Who can open it

Role	What they can do
Admin	Everything on this page: create, import, edit, duplicate, export, delete, preview, run and publish agents; manage skills and collaborators.
Builder	The same as an Admin. Creating an assistant’s share link at publish time takes the Admin role.
Auditor	Open the list and every agent’s page. No actions, no preview and no runs.
Reviewer, End user	Cannot open it. A Reviewer answers the approvals agents raise in Approvals.
Platform admin	Everything an Admin can, in any tenant. A tenant picker appears at the top of the page.

What you see

The Agents list

The header: the title Agents and, when you can see more than one tenant, a tenant picker.

Activity, last 7 days: four figures for the tenant’s agents together.

Figure	What it counts
Runs, last 7 days	Agent runs started in the window.
Completed	The share of those runs that finished, with the count (for example `12 of 14`).
Tokens, last 7 days	Tokens the runs spent.
Approvals waiting	Tool calls held for a person; the figure links to Approvals.

The figures are counted from the newest 500 runs; when there are more, the strip says so.

The table, with a search box (it matches name, description and model):

Column	What it shows
Name	The agent’s name, linking to its page. A Platform tag marks an agent that ships with the platform.
When it is used	The routing key: the description other agents read to decide whether to hand work to this one. Not used for routing when it is empty.
Model	The chat model the agent answers with.
Tools	How many tools the agent is granted.
Skills	How many skills are attached.
Collaborators	How many agents it may delegate to.
Published as	The assistants that publish this agent, linking to each; Internal when none does.
Updated	When it last changed.
Actions	Admin and Builder only: Preview, and a menu with Duplicate, Export YAML, Publish as assistant and Delete.

The toolbar (Admin and Builder): Import YAML and Create agent.

When the tenant has no agents, the table reads No agents yet, with Create agent and Import YAML for the roles that may use them. When the list cannot be read, the table shows the reason and a way to try again.

An agent’s page

The header carries the agent’s name and routing key, a status (Published or Internal) and, for an Admin or Builder, Preview, then Publish as assistant (or View assistant once an assistant publishes it), and a menu with Edit, Export YAML, Duplicate and Delete.

Below it, in order:

Activity, last 7 days: the same four figures, for this agent only.
Shipped with the platform: a notice on a platform-shipped agent, which cannot be edited (see Platform-shipped agents).
How it is connected: a graph of the agent and everything around it: Reached through (the assistants that publish it, or a note that it has no link of its own), Tools it can call (tools that need approval are marked Needs approval), Skills it can use, Knowledge it reads and Agents it hands work to. A dashed line is knowledge it reads; every other line is something it can use or the way it is reached.
Identity: Name, When it is used, Reachable from outside (No — only a published assistant has a URL, or the assistants it is reached through), Updated, and Runs, a link to Runs of this agent.
Behavior: Model, Instructions, Temperature.
Knowledge: Collections (each links to its collection) and Chunks per answer.
Tools & skills: the granted tools grouped by server, each marked Allowed or Needs approval; the attached skills with what each narrows the tools to; and Effective tools, the arithmetic of what the agent may call while each skill is in use (see Skills).
Collaborators: the agents it may delegate to, with Manage for an Admin or Builder (see Collaborators).
Run: a message box and Run, for an Admin or Builder (see Preview and test).
Limits & approvals: Tool calls per run, Model calls per run, Tokens per run (each shows (platform default) when unset) and Cost at the ceiling.

When the agent does not exist, or belongs to another tenant, the page reads That agent is not here, with All agents to go back.

What you can do

Export an agent as YAML

On the list, open the row’s menu and choose Export YAML; on an agent’s page, open the menu and choose Export YAML.
Olympusx AI copies the agent’s file to your clipboard and confirms with YAML copied.

The file names what the agent links to (servers, skills, collaborators and collections) by name, never by internal id, and it carries no credential. Review it in a merge request and import it into another tenant or another installation. The file’s format is described in Create and edit an agent.

Duplicate an agent

Choose Duplicate from the row’s menu or the agent page’s menu.
In the dialog, keep or change the Name. It starts as the original’s name followed by (copy).
Choose Duplicate.

The copy carries the instructions, knowledge, tools, skills and collaborators, and it is yours to edit, whoever shipped the original. The copy’s model must be a chat model this tenant has deployed.

Delete an agent

Choose Delete from the row’s menu or the agent page’s menu.
Confirm with Delete agent.

Deleting cannot be undone. An agent that an assistant publishes cannot be deleted: the dialog names the assistants, and Olympusx AI refuses until each one is pointed at another agent or removed.

Publish an agent as an assistant

Choose Publish as assistant on the agent’s page, in the list row’s menu, or on the Preview page.
Complete the assistant form, which starts from this agent. Publish and share describes every field.

Once published, the header shows View assistant instead, and the Published as column names the assistant.

Platform-shipped agents

An agent marked Platform ships with Olympusx AI. It cannot be edited, so what it does is always what the platform shipped: Edit is unavailable, and Delete is available only to a platform admin. To change one, Duplicate it and edit the copy.

Limits

Limit	Value
Agent name	1 to 80 characters, unique in the tenant.
Routing key (When it is used)	At most 280 characters, counted as characters in any script.
Activity figures	The last 7 days, counted from the newest 500 runs.
Delete	Refused while an assistant publishes the agent.