Agents

Agents are your AI employees. On this screen you create them and set their personality, behavior, model, skills and what they are allowed to do. The screen is scoped to a workspace. Agents are shown as tiles: icon, name, model (or “default model”), status, a preview of the system prompt, and “Profile” / “Delete” buttons.

Create your first AI employee in 3 steps

If you are doing this for the first time, three steps are enough — fine-tuning can wait:

Open the “Agents” screen and click “+ Create”.
Set a name (e.g. “Support Assistant”) and a system prompt — a short description of the role: who this employee is, what it helps with and how it talks.
Click “Save”. You can leave the model empty — the default from the provider chain is then used.

What you’ll see: the new agent appears as a tile with a system-prompt preview. You can test it right away without connecting a channel: on the “Sessions” page click “+ New session” and message it.

[!TIP] Until a real model is connected, replies come from a deterministic mock — enough to confirm the agent is created and answering. Connect a provider (OpenAI, Anthropic, YandexGPT, GigaChat, Codex and more) on the Models page.

Creating an agent

The “+ Create” button opens a modal with the basic fields:

Name — required, shown in lists and channels.
Avatar — upload or pick a photo (an image).
System prompt — the text that defines the role and behavior rules.
Model — a model id; leave it empty to use the default from the provider chain.

That is enough for the agent to work. Fine-tuning happens in the profile.

Agent profile

The “Profile” button opens an editor with independently saved blocks.

Identity

Short and full name, role, how to address the user, mood, gender (unset / male / female / other), character (free text), personal phone and e-mail. These fields form the employee’s “personality” and how it introduces itself in channels.

Mailbox

Appears once the agent has a personal e-mail: IMAP and SMTP parameters (host, port, user, password). Passwords are write-only and never read back. This lets an agent work with its own inbox.

Behavior sections

A set of accordions, each with an “enable” checkbox and a text body: Known user, Rules, Safety, Capabilities, Memory policy, Refusal style, Response format, Escalation and Persona constraints. This — not a separate set of “permission checkboxes” — is where you define what an agent may and may not do: describe it in Safety, Capabilities and Rules.

Shared blocks

You can link company-wide shared blocks from the workspace catalog (a single set of company rules) — see Skills & prompts. A block shows a sensitivity badge.

Skills

The agent’s own skills (create/edit/delete: name, description, body, enabled, order; the slug is generated by the server) plus linking the workspace’s shared skills.

Setting rights and model

Rights/permissions are described in the Safety, Capabilities and Rules sections and through the set of available skills and tools. Real risky actions pass through Approvals and Moderation.
Model can be set at creation (free text) or left empty for the default; which models are allowed at all is on Instructions → allowed models.

If it didn’t work

The agent replies in a generic/templated way — the mock model is probably still active. Connect a real provider on the Models page.
I don’t see the agent in a channel — creating an agent isn’t enough; you must bind it to the channel in the “Bindings” block on the Channels screen.
I set a model but the agent errors out — that model may not be in the allowed list, or the provider’s Test didn’t pass; check Models and allowed models.
I want to limit what the agent may do — this is set not via “permission checkboxes” but in the Safety, Capabilities and Rules sections of the agent profile, plus through Approvals.

Tips

Keep the system prompt short and push details into sections and shared blocks — they are easier to reuse across agents.
Sensitive shared blocks are edited through the approvals queue.

Orchestration — wire several agents into a graph.
Skills & prompts — reusable blocks.
Models — providers and priority.