Agents

Reusable AI agents with their own role, knowledge, tools and model – private or shared with people, teams, projects or the whole company.

Agents (formerly “assistants”) are persistent chat personas. Instead of retyping the same instructions in every conversation, you define role, approach, knowledge, tools and model once – and can then start a chat that works exactly that way at any time.

What an agent is

An agent bundles several things into one reusable profile:

a persona (system prompt) that steers behaviour in every chat
a preferred model and model parameters (creativity, tokens, reasoning)
its own knowledge (attached documents) used automatically when answering
tools via integrations (e.g. web search, OneDrive, Google Drive)
conversation starters and a greeting message for a smooth start
optional subagents and skills for more complex tasks

New agents are private: only you see and use them until you release them via Share – to individual people, a team, a project or the whole company. Owners and admins see other people’s private agents only by name (to delete them), not their instructions.

The Library

You reach agents via the Agents entry in the sidebar. The page has three tabs:

Agents – every agent you may use
Workflows – multi-step, form-based flows (a separate module)
Templates – ready-made templates to copy

The Library Agents tab: search, filter and launch a chat with one click.

Searching and filtering

In the Agents tab, the toolbar offers:

a search box (matches name and description)
the filters All, Mine (created by you), Shared with me, Company (shared company-wide) and Pinned
a label bar: click one or more labels to narrow the list

Highlighted agents automatically float to the top of the list.

An agent card

Each card shows the icon (emoji on the brand colour), the name, a two-line description, up to three labels, and in the footer the model plus a Web badge if web search is on. A shield icon marks verified agents.

Click the card → immediately starts a new chat with this agent.
Pencil icon → opens the editor. You only see it if you may edit the agent: as its creator, with the “Edit” share, or as an owner or admin for agents you may use.
Trash icon → deletes the agent (creator as well as owners and admins).
Pills show which team or project the agent is assigned to; a click opens the agents list there.

Creating an agent

Open Agents in the sidebar.
Click + New agent in the top right. A blank agent ("New agent") is created immediately and the editor opens.
Fill in the sections (see below) and save.
Click Publish so your configuration takes effect in real chats. Only the creator, owners and admins may publish.

Start faster from a template: switch to the Templates tab, filter by category and click a suitable template. In the preview, choose Adopt and customize (opens the editor with pre-filled values) or Adopt and chat (starts a chat right away). The copy belongs to you and is private at first.

The editor

The editor is split in two: configuration on the left, a live preview on the right (on large screens) where you can try the agent out instantly. The header bar offers the actions ▶ Test, Clone, Save and Publish, plus an Unpublished changes indicator when relevant.

Identity

Avatar: upload a custom image via Image (PNG, JPEG or WebP, max. 2 MB). An uploaded image takes precedence over the emoji; Remove resets it.
Emoji: a short symbol (up to 4 characters) shown on the brand colour.
Colour: the brand colour (hex #RRGGBB) for the icon background and card.
Name (required, max. 120 chars) and Description (max. 500 chars).

Behaviour

System prompt (required, max. 20,000 chars): the core instruction that defines the agent's behaviour in every chat – role, tone, rules, boundaries.
Optimize: lets the language model rewrite your prompt (sharper role, clear rules, same language). Undo restores the original.
Default model: the preferred model for chats with this agent. — means the normal model picker applies.
Web search: lets the agent research the web when needed (on by default).
Research the web on every answer (off by default): automatically runs a web search before every answer and feeds in the results. More reliable than leaving it to the model – but only sensible for pure research agents.
Integrations: bind providers (e.g. Outlook, Google Drive, Jira) to the agent. Their tools are enabled automatically in chats with this agent. The user still needs their own OAuth connection to the provider.
Creativity (temperature, 0.0–1.0): how creative the model may be. Recommendation: 0.7 for agents without attached knowledge, 0.3 with attached knowledge.
Advanced model parameters (collapsible): Max tokens (maximum answer length, 1–32,000) and Reasoning effort (low/medium/high, only relevant for reasoning models).

Input type: prompt or form

Prompt (standard): a normal chat – the user types freely.
Form: instead of a free-text chat, the user fills in predefined fields. Define them in the Fields section. Per field you set a key, label, description, whether it is required, and the type: text, longtext, select, number, email, checkbox, date or file (max. 20 fields).

Greeting message

An optional text (max. 2,000 chars) that the agent automatically "says" as the first message of every new chat. Ideal for setting context or asking up front for the information you need.

Conversation starters

Up to four pre-written opening prompts (max. 500 chars each). They appear as clickable cards before the user types their first message – one click drops the text straight into the input.

Knowledge

Attach documents the agent uses automatically when answering (RAG). There are two sources:

Link company-wide documents – references existing company-wide documents.
Own documents – uploaded directly for this agent via Upload file or Pick from cloud (OneDrive, SharePoint, Google Drive).

In chats with this agent, knowledge retrieval uses exactly these documents – linked and own – plus files attached in the chat. Other company-wide or project documents are not added automatically.

Knowledge vs. attachments: the knowledge stored here is permanently available in every chat with the agent. It does not replace attaching individual files directly inside a chat.

Skills

Attach skills that are active only within this agent. The agent can call them itself via a tool when needed. Attach skill opens a searchable picker of the active skills.

Subagents

Link other agents in your company as subagents to which this agent can delegate sub-tasks. Each subagent works independently (its own knowledge, possibly web search) and returns its result.

Delegation is deliberately bounded: at most two levels deep and with a call budget per answer. An agent cannot be its own subagent, and direct cycles (A→B while B→A already exists) are rejected.

Actions

If the agent uses integrations (e.g. Outlook, OneDrive, Jira), set under Actions what it may do per tool. Tools are grouped into Read, Change and Send:

Setting	Effect
Allowed	The agent runs the tool directly.
Ask first	A card with the details appears in the chat; only your Run triggers the action, Reject discards it.
Off	The tool is not available to the agent.

Default: reading is allowed, changing and sending ask first. Exception: SQL queries on connected databases also ask first, even though they only read. If several agents share a chat, the strictest setting applies.

Only owners and admins may set Allowed for sending, for changes that reach beyond your own account (e.g. in shared storage, spreadsheets or tickets) and for SQL queries. This only takes effect while the agent’s creator is an owner or admin. If someone without that role later changes the instructions, integrations or actions, such settings fall back to the default.

Sources

Under Sources you tighten your company’s source policy for this agent: priority (as company default, company knowledge first, company knowledge only) plus additionally blocked or exclusively allowed websites. A stricter company setting always applies.

Use Share at the top of the editor to decide who uses the agent:

Visibility: Private (default – only you and the shares) or Whole company (everyone can use it).
Shares to a person, a team or a project, each with the role Use or Edit. A team reaches all members, a project its members and its home team.
Anyone may share with an individual person. Shares with teams, projects and the whole company require the “Share assistants with teams, projects and the whole company” permission (admins only by default, see Roles and permissions).
Home project: with a home project only members of that project use the agent.

When an agent is shared more widely, all linked documents must cover the new audience – otherwise Kasimir rejects the share and says why. This way no documents reach people via an agent who may not see them.

Labels

Add up to eight labels (max. 30 chars each). They help others find the agent in the Library via the label bar.

Usage

A read-only overview: number of Messages, Tokens consumed, estimated Cost (in EUR) and Last used.

Verification (admins / Kasimir staff only)

Mark as verified: adds a verification badge (shield icon) to the card.
Highlight in library: makes the agent appear at the very top of the list.

Draft and publishing

This is the most important concept when editing an agent.

While editing you change a draft. Save stores the draft. Once the agent has been published, saving does not change behaviour in running or new chats; before the first publish, chats use the saved state directly.
After that, chats always use the most recently published version of the behaviour. Only Publish applies the current state (it saves automatically first).
Whenever the draft and the published version differ, the header shows Unpublished changes.

Draft, save, publish: chats always use the published version.

Which fields get "frozen"? Only behavioural fields are published (snapshotted): system prompt, default model, web search, forced web search, creativity/tokens/reasoning, conversation starters, input type and form fields, plus the settings under Sources (priority, blocked and allowed websites). Display fields like name, emoji, brand colour and avatar are always live immediately – they don't need publishing.

Test and clone

▶ Test: saves the current state and immediately opens a fresh chat with the agent. As long as the agent has never been published, you test exactly what's in the form; after that, this chat – like the live preview – uses the most recently published version. So you only see behaviour changes there after Publish.
Clone: creates a private copy that belongs to you (name suffixed with "(Kopie)"). The configuration is copied; linked documents only as far as they may be shared with the copy. The agent's own file uploads, its avatar image, the settings under Actions, verification and highlighting, and the published version are not copied – the copy starts unpublished.

Using an agent in chat

Launch a new chat bound to the agent either from the Library (click a card) or from a pinned entry in the sidebar. In project and team chats, only agents that are company-wide or shared with that space can be used.
On start, the greeting message and conversation starters appear first (if configured).
During the conversation the agent's persona, preferred model, attached knowledge and bound integrations apply automatically.
You can also address an agent per message via an @-mention – then only that agent answers that specific message.

Limits and notes

Length limits: name 120, description 500, system prompt 20,000, greeting 2,000 chars; up to 4 conversation starters (500 each), 8 labels (30 each), 20 form fields.
Avatar images may be at most 2 MB.
Bound integrations enable the tools automatically but do not replace the user's own OAuth connection to the provider.
Delete, Publish and managing shares are reserved for the creator and admins/owners; verification is reserved for admins and Kasimir staff.
Deletion is permanent and cannot be undone.

BackWork in a folder NextWorkflows