Workforce

Workforce is the workspace section where digital employees live: who they are (roles), what they do (scenarios), how much they spend (budgets) and which machines they run on (nodes).

Teams and roles

A role is a document: what the agent does, what is forbidden, which channels it may use and its risk class.

Class	What is allowed
r0	reading public sources only
r1	working with data inside the organization, never writing outside
r2	writing to people; every message is approved by a human

Roles are grouped into teams: sets of roles and scenarios that together cover a piece of work. Every organization already has 7 teams from 47 roles (catalog, in Russian). A role can have a manager: a stuck task, an uncertain answer or a conflict with the rules moves up the chain of managers or to a human, as the role’s escalation policy decides. Open escalations are in the Escalations card: “Take” stops the climb up the chain, “Close” means the question is settled.

Role versions and rollback

Roles and scenarios are versioned. Agents get only published versions; drafts never reach them. In the Roles card: team → role → versions with Publish, Unpublish and Return vN…. Before each action the panel says which version agents will get afterwards and lists published scenarios that reference the version being unpublished: they keep working with the pinned version, but their new versions will not publish with it.

“Name and description…” on a role creates a draft of a new version, a copy of the one agents have, with the new name and description; agents get it after Publish.

Scenarios

A scenario is the sequence of steps an agent follows for a task: actions, waiting for input, compensations on failure, conversation phases for channels. A scenario references a role by version number, so rolling back a role does not change an already published scenario. The format is described in the scenario reference (in Russian).

A task on a scenario is taken by an agent that has all the needed connectors: work never goes to a machine without, say, WhatsApp. The agent asks for work with long polling: the cloud holds the request for up to 30 seconds and answers as soon as a task appears. Along the way the agent writes checkpoints; the run, what it was started with, what the agent asked the model and what it answered, is visible on the task, with keys and tokens masked.

Budgets and pause

• The task budget is enforced by the server: once it is exhausted, no new leases on the task are handed out.
• The monthly agent budget in dollars: 80 % is a warning, 100 % is a budget pause. It lifts itself in the new month or when the budget is raised.
• An operator pause holds until you lift it: started tasks finish, new ones are not handed out. A disabled agent cannot re-enroll on its own.
• A budget raise request appears in Approvals automatically when the budget is exhausted; approval returns the task to the queue.

Spend is counted per model call, attributed to the agent and the task; model prices are first-party, the cache discount is not applied.

Nodes and enrollment keys

A node is one HolyAgent on one machine. Enrollment is by a one-time key from the Nodes card: set the agent name and a validity period (72 hours by default); the key is shown once and works only for that name. After the first run the agent receives a permanent key and keeps it; an unused one-time key can be revoked in the same card. The card shows the status, version and manager of every node.

An agent can ask for a new employee itself, from its panel, with a role and a reason. The request appears in Approvals; approval issues a one-time permission to enroll, a rejected or expired (7 days) request issues nothing.

Next

• HolyAgent: how to install the agent and connect channels.
• Company package (in Russian): how to move roles and scenarios as a file.