# agents

> Reference for the agents tool of Agent host: list, get, send, approvals, digest and usage, with every action.

Source: https://agenthost.geenius-os.tech/docs/reference/agents/

The hosted agents of one workspace: see them, talk to them, read their digests and approvals queue and check usage. Each agent keeps its own memory and asks a person before anything risky. The door can watch and talk; it cannot decide.

Call it as `geenius_agenthost(tool: "agents", action: "<action>", args: {...})` through Geenius OS.

## Actions

| Action | What it does |
|---|---|
| `list` | The workspace's agents with their state (draft, running, paused or error), folder, tags, last run and last error. Filter by text, status, tag or archived. |
| `get` | One agent: its state, versions, connectors (never a secret), the newest definition and its last digest. |
| `send` | Sends one message to a deployed agent and returns its answer. A risky action the agent wants to take does not happen: the answer is a pending approval id and a person decides in the app. |
| `approvals` | The requests waiting for a person of the workspace: what, which agent, since when. You cannot decide them here. |
| `digest` | The last result of a scheduled agent: its digest text and when, or null. |
| `usage` | What the plan holds and what is used: agents, this month's AI credits, tool calls a day. With an agent id, that agent's usage per day. |

> [!TIP]
> Ask for the exact arguments of any action with `action: "help", of: "list"`. The server answers from its own definitions, so that answer is always current.

> [!NOTE]
> There is no approve action. A risky call an agent makes through send comes back as pending approval, and only a person of the workspace answers it in the web app. Otherwise an AI could approve its own webhook, which is what the approval gate exists to prevent.