AI Agent NEW

Kanata AI Agent

Manage AI agents with a harness

Kanata AI Agent carries out work requested in chat, based on the internal knowledge in the Library. You can entrust it with everything from answering questions to creating documents, operating apps, creating images and drafting slide outlines. As a harness that keeps AI running safely, you set reference folders, execution permissions and the autonomy level in advance.

The AI agent chat screen, where a sales support agent proposes a plan to create an image

Kanata AI Agent overview

Kanata AI Agent runs on Kanata's servers and, based on Library content, uses apps and Studio as tools to carry out research and work on behalf of the requesting user. Its actions are limited to that user's permissions.

01

What it covers

Answering questions in chat, creating and editing documents, operating apps such as scheduling recordings, creating images and drafting slide outlines

02

Best suited for

Organizations with many internal inquiries and routine tasks that want to expand AI use while managing what is entrusted to AI and tracking its usage

03

What you can aim for

Reducing the effort of handling inquiries and creating documents, and making it possible to check what AI based its answers on and what it did

Problem

Challenges organizations face in using AI agents

Surveys show that only some organizations in Japan connect generative AI to internal data. Concerns about generative AI include the accuracy of information, information leaks and accountability when problems occur, and more than half of companies report a shortage of full-time employees.

Share with an environment where generative AI learns from or refers to internal data (Japan)

24.5%

The share answering "we have generative AI learn from internal data or have built a database generative AI can refer to". It was 57.2% in the US, 54.4% in Germany and 73.0% in China.

Source: Ministry of Internal Affairs and Communications, "White Paper on Information and Communications in Japan 2026", Figure I-2-1-5 (survey conducted January–February 2026; in Japan, 515 managers or above at companies)

Share of companies citing "accuracy of information" as a concern or issue with generative AI

50.4%

"Risk of information leaks" was 33.5%, and "establishing rules such as accountability when problems occur" was 25.5% (up to three answers allowed).

Source: Teikoku Databank, "Survey on Corporate Trends Regarding Generative AI (March 2026)" (survey conducted March 17–31, 2026; 10,312 valid responses)

Share of companies reporting a shortage of full-time employees

51.3%

As of July 2026. 28.8% of companies reported a shortage of non-regular employees.

Source: Teikoku Databank, "Survey on Corporate Trends Regarding Labor Shortages (July 2026)" (survey conducted July 17–31, 2026; 10,518 valid responses)

Challenges on the ground

Even with generative AI chat in place, it may not refer to internal documents, and "we can't tell what answers are based on"

We want to entrust work to AI, but "we don't know what it will do" before it runs

Each department uses separate AI tools, and "knowledge and work aren't connected"

Short-staffed, people "lose time" to routine tasks and inquiries

Simply rolling out a chat tool makes it hard for internal knowledge to become the basis for AI, and leaves few records of work entrusted to AI. You need a mechanism that answers based on internal knowledge and reviews the plan before acting.

Scenes

How Kanata AI Agent is used

Examples of how departments use Kanata AI Agent. The video shows how a department admin sets what is entrusted to AI and how it is reviewed.

The companies and people shown are fictitious.

Scene 01

Inquiries about internal policies

HR and admin staff

BeforeHandling inquiries about internal policies took up a lot of time.

  1. Set reference foldersSet the folder where internal policies are stored as a folder the AI agent refers to.
  2. Ask in a DMWhen employees ask in a DM, the AI agent gives a first-line answer based on internal policies.
  3. Answers within viewing permissionsDocuments the person asking cannot view are not used as a basis.

With KanataFirst-line answers based on policies are left to the AI agent, and staff handle consultations that need judgment.

Challenges solved: we can't tell what answers are based on, lose time

Scene 02

Turn meeting decisions into tasks

Sales staff

BeforeThey wanted to entrust work to AI but didn't know what it would do before it ran.

  1. Request in a DMIn a DM with the sales support agent, ask it to turn the sales meeting decisions into tasks for each owner.
  2. Review the planAt autonomy level 2, the requesting user reviews and approves the plan, including writes and app executions.
  3. Run the approved scopeOnly the approved scope is run. If an action beyond the approved scope becomes necessary, it stops at that point and asks again.

With KanataYou review the plan before execution and entrust only the approved scope to the AI agent.

Challenge solved: we don't know what it will do

Scene 03

Setting what is entrusted to AI

Department admin

BeforeEach department used separate AI tools, and knowledge and work weren't connected.

  1. Set what it handlesSet custom instructions, reference folders, execution permissions, autonomy level and the models to use.
  2. Agents for each departmentCreate agents for each department or purpose, such as an admin help desk or sales support, and share them in the project.
  3. Check usageSpace admins check the remaining credits and purchase additional credits as needed.

With KanataInstead of banning AI, define what is entrusted to it and use it in departmental work.

Challenge solved: knowledge and work aren't connected

These scenes are examples of use. The people shown are fictitious.

Before and after

Compared with before AI agents

Until now, people carried out work by operating each system's screens, and generative AI chat handled answers and drafts. Kanata AI Agent carries out work by operating Kanata's features directly.

PointBefore generative AI (manual operation)Generative AI chat (for answers)Kanata AI Agent
How work gets doneStaff open and operate each system's screensAI creates answers and drafts; people mostly handle input and operations in systemsThe AI agent presents a plan and operates Library, Apps and Studio features directly. It keeps running even if you close the screen after requesting
Information used as the basisStaff look for documents and judge with their own knowledgeMainly what is entered in the conversation. Internal documents may need to be entered each timeBased on internal knowledge in the reference folders
PermissionsSystem permissions are granted to each userPermissions may be managed separately for the AI tool and internal systemsAI has no permissions of its own and acts within the requesting user's permissions
How to requestAsk the person in charge verbally or by emailOpen a separate AI tool and enter the necessary informationRequest in a chat DM, or with the button at the bottom right without leaving the screen you are working on
RecordRemains as each person's workMay remain as conversation historyCredits are counted as usage by the requesting user

Server side

How it operates Kanata on the server side

Kanata AI Agent does not operate the user's PC or browser screen; it uses Library, Apps and Studio features as tools on Kanata's servers. Actions are checked against the requesting user's permissions.

01 · Request

Requests from chat and the bottom-right button

  • Requests in chat DMs
  • Requests from the AI agent at the bottom right (using the current screen as context)
  • [Ask AI] in minutes and e-Learning materials

02 · Kanata AI Agent

Plan and permission checks

  • Checked against the requesting user's permissions
  • Plan reviewed according to the autonomy level
  • The requesting user approves the plan

03 · Tools

Operating Kanata features

  • Reading and writing in the Library
  • Operating apps (scheduling meetings, etc.)
  • Creating images and outlines in Studio
  • Web search (if allowed)

04 · Output and records

Answers, saving and records

  • Answers in DMs
  • Saving to the Library
  • The requesting user's credits are consumed
  • Unlike approaches that record and replay screen operations, it uses Kanata's features directly, so it is less affected by changes to screen layout.
  • Execution continues even if you close the screen after requesting. Actions that need confirmation pause as "awaiting confirmation" and appear in the list of items awaiting confirmation.

Platform

The advantages of being part of an AI platform for organizations

Kanata AI Agent runs on the same foundation as the Library, Apps, Studio and chat. There is no need to set up connections to internal knowledge, permission checks or usage records separately for AI.

Internal knowledge in the same Library

Minutes, documents and app data accumulate in the same project's Library. Choose reference folders, and they become the basis for the AI agent.

Same permission checks as users

AI has no permissions of its own and acts within the permissions the requesting user has in the Library, Apps and Studio.

Operating Apps and Studio directly

App operations such as scheduling meetings, and creating images and slide outlines, are carried out as Kanata features.

Credit management

Credits are counted as usage by the requesting user, and space admins check the remaining credits and purchase additional credits as needed.

Managing consumption with credits

The AI agent's credits are consumed as usage by the requesting user.

Requests from the bottom right of the screen

Use the button at the bottom right of the project screen to ask the AI agent without leaving the screen you are working on. The screen being displayed is used as context.

Features

Kanata AI Agent features

Kanata AI Agent takes requests from chat DMs or the button at the bottom right of the screen, and answers and acts based on internal knowledge in the Library. Custom instructions, reference folders, execution permissions and autonomy level are set for each agent.

Custom instructions and reference folders

Describe the role, answering style and rules to follow in a single text, and set the folders it refers to as a basis.

Execution permissions and autonomy level

Set read, write and execute permissions for each tool, and the scope for plan review.

Requests from chat and the bottom-right button

Switch between multiple chats for each agent, and make requests from the screen you are working on.

Main features

Settings

Custom instructions and reference folders

Set the answering style and scope of sources for each agent

Set the name, icon and custom instructions. In custom instructions, describe the role, answering style and rules to follow (such as what it should not answer) in a single text (up to 8,000 characters). Choose reference folders, and all content in them becomes candidate sources, so there is no need to select files one by one.

The settings screen of a training assistant, setting its name, icon, custom instructions and reference folders
Approval

Plan review and autonomy levels

Review the plan before writes and app executions, according to the autonomy level

Choose from autonomy level 1 (suggest only), level 2 (review the plan and run) or level 3 (run automatically). At level 2, only the approved scope runs continuously, and it stops and asks again when an action beyond that scope is needed. Publishing, overwriting, deleting, schema changes and Studio runs that include four or more creations are confirmed even at level 3.

A sales support agent presenting an execution plan and waiting for approval
Chat

Requests in chat

Use multiple chats for each agent

In chat DMs, create multiple chats with each agent and use them for different topics. A chat is titled from the first request and can be renamed later. Documents the user in the conversation cannot view are not used.

A one-to-one chat with the HR assistant, drafting a paid leave request
Anywhere

The AI agent at the bottom right

Ask right from the screen you are working on

Use the button at the bottom right of the project screen to ask the default AI agent. The current screen is provided as context, and you can remove the context with [Remove]. The conversation remains in your chat DMs, and [Open in DM] lets you continue there. Plan review, permissions and credit checks work the same as in DMs. The button is not shown on chat DM screens or full-screen views (such as the AppStudio canvas).

The list of AI agents, showing a sales support agent, an HR assistant and a training assistant with their autonomy levels

Approval

Types of actions and when they are confirmed

The agent's actions fall into seven types, and each autonomy level determines whether they are confirmed before running. Read-only requests are answered directly at any autonomy level.

CategoryTaskLevel 1
Suggest only
Level 2
Review plan and run
Level 3
Run automatically
ReadRead documents and recordsNo confirmationNo confirmationNo confirmation
WriteCreate and edit documents in the LibrarySuggest onlyConfirmNo confirmation
Run appOperate apps (scheduling meetings, etc.)Suggest onlyConfirmNo confirmation
Run StudioCreate images, slide outlines and moreSuggest onlyConfirmNo confirmation
PublishPublish (apps are published by users in AppStudio)Suggest onlyConfirmConfirm
Schema changeChange content type fieldsSuggest onlyConfirmConfirm
Delete and overwriteDelete or overwrite (adds a version)Suggest onlyConfirmConfirm

Studio runs are confirmed even at autonomy level 3 when a single plan includes four or more creations.

Only project admins can choose autonomy level 3. Only the requesting user can approve a plan.

Rules

Rules for using the AI agent

Rules on permissions, approval, consumption and conversations when making requests to the AI agent.

  1. Within the requesting user's permissions
    Actions always run on behalf of the requesting user and within that user's permissions.
  2. Only documents the user can view
    Only documents the user in the conversation can view are used as a basis.
  3. Approval by the requesting user
    Only the requesting user can approve a plan.
  4. Consumed as the requesting user's usage
    Credits are consumed as usage by the requesting user.
  5. How conversations are handled
    Conversations remain in chat DMs and are not saved to the Library.
  6. Default agent
    You can set a default AI agent for each project, which is what the button at the bottom right of the screen talks to. The button is not shown on chat DM screens or full-screen views (such as the AppStudio canvas).
  7. Credit consumption and management
    Credits used by the AI agent are counted as usage by the requesting user. Space admins can check the remaining credits and purchase additional credits as needed.

Column

Related columns

FAQ

FAQ

QWhat happens to the existing AI Chat?

AI Chat will be discontinued in the new version. You will ask AI for help through chat DMs and the AI agent at the bottom right of the screen. AI Chat apps, settings and conversations will not be migrated to the AI agent or chat. How past conversations are handled (the period for viewing and exporting them) will be announced in our news. Save documents you want the agent to refer to in its reference folders.

QHow is the scope the agent refers to determined?

Within the "reference folders" set for each agent, only what the user in the conversation can view is used as a basis. When writing, only folders the requesting user can write to are eligible.

QCan we use multiple chats with one agent?

Yes. Create multiple chats with each agent and switch between them for different topics. A chat is titled from the first request and can be renamed or deleted. Documents created remain even if the chat is deleted.

QCan we ask the AI agent from the screen we're working on?

Yes. Use the button at the bottom right of the project screen to ask the default AI agent. The current screen is provided as context, and the conversation remains in your chat DMs. The button is not shown on chat DM screens or full-screen views (such as the AppStudio canvas).

QCan we choose the AI model to use?

Yes. For each agent, set the model selected first in chat and the models available. In the DM input field, the requesting user chooses a model from among them.

QDoes the AI agent keep running after the screen is closed?

Yes. Because the AI agent runs on Kanata's servers, execution continues even if you close the screen after requesting. Actions that need confirmation pause as "awaiting confirmation" and appear in the list of items awaiting confirmation.

QCan it post to Slack or Teams?

Posting to external services is not supported. If asked to post, it only provides a draft of the text and does not send anything externally.

QWhose credits does the AI agent use?

Credits are consumed as usage by the user who made the request. Space administrators can check the remaining credits and purchase more as needed.

Last updated: