How to create custom AI agents in Vaiz to automate routine work

Create your own AI agent to take care of routine work in Vaiz. Vaiz AI agents follow custom instructions and can run on a schedule, in response to events, or manually. You choose the AI model, scope, and allowed actions. In Ask mode, proposed changes require a reviewer’s approval before they are applied.

Set up an agent manually, or describe what you want it to do and let Vaiz configure it based on the context of your work.

Step 1: Open the agent creation window

Open the agent creation window in Vaiz and choose how you want to set up your agent:

• Manually: define the instructions, triggers, AI model, and other settings yourself.
• From a description: describe what you want the agent to do, and Vaiz will configure it based on the context of your work.

If you choose setup from a description, explain the agent’s responsibility, where it should work, and when it should run. For example:

Create an agent that reviews open tasks in this project every day at 09:00 UTC. It should prepare a short summary of overdue work and reported blockers, with references to the relevant tasks.

For manual setup, follow the steps below.

Step 2: Write the agent’s instructions

In the Instructions field, describe what the agent should do. Start with one clear responsibility, such as summarizing task activity or checking task descriptions.

Structure your instructions around four points:

• Goal: what the agent should accomplish.
• What to review: which information and time period it should cover.
• Rules: what it should include, avoid, or do when information is missing.
• Output: how the result should be presented.

For example, a Summarizer agent can explain what has changed in a task:

Goal: When asked, briefly summarize the recent history of the task.
What to review: Task activity and comments from the last seven days. If there are fewer than three events, use the ten most recent events regardless of date. Include changes to status, assignee, priority, and due dates, along with completion updates and reported blockers.
Rules: Summarize only what is recorded in the history. Do not infer or invent information. Combine minor changes of the same type and state who made the changes and when. If the history is unavailable or empty, say so.
Output: Return a short, clear summary, with the most important updates first.

You can use @ to mention relevant items in your instructions.

Click Continue.

Step 3: Set the agent’s behavior and scope

Choose how the agent handles changes, when it runs, and where it works.

Mode

The Mode setting controls how proposed changes are handled.

In Ask mode, the agent’s proposed changes wait for a reviewer’s approval before they are applied. This lets you review suggestions before they affect your work.

When the agent runs

Vaiz AI agents support scheduled, event-driven, and manual runs.

For a scheduled run, select On schedule under Trigger. Set the Interval and Hour (UTC). For example, choose Every day at 09:00 UTC for a daily project briefing. The time shown is in UTC, so account for your local time zone.

For an event-driven run, select an available event trigger that matches your workflow.

You can also run the agent manually when you need a result or want to test its instructions.

Scope

Use Project and Boards to define where the agent works. Choose the whole space or narrow its scope to the relevant project and boards.

Match the scope to the instructions. An agent preparing a project summary should work with the project that the summary covers.

Click Continue.

Step 4: Choose allowed actions and an AI model

Under Actions, select what the agent is allowed to do:

• Edit tasks
• Create tasks
• Manage milestones
• Write comments
• Create & edit docs
• Write Space Memory

Select actions individually or use Allow everything to enable all listed actions.

Match the permissions to the agent’s responsibility. For example, an agent that leaves feedback on task descriptions needs permission to write comments.

Under Model, choose the AI model that will carry out the work. Test it with a representative task to check whether it follows your instructions and produces a useful result.

Click Continue.

Step 5: Set access and integrations

Choose who reviews proposed changes, whose permissions the agent uses, and whether results should be posted to Slack.

Reviewers

Under Reviewers, select the managers who will approve proposed changes in Ask mode. Reviewers are managers of the selected scope.

Execution owner

Under Execution owner, select the manager whose access the agent will use.

Every scheduled, event-driven, or manual run uses this manager’s permissions to read information and make changes. The agent can only access what that manager can access, within the selected scope.

For more about access in Vaiz, see How to set roles and permissions.

Slack results

To post run results to Slack, choose a channel under Post results to a Slack channel.

First, connect Slack to your space in App Center and allow the channels you want to use. Then select a destination channel in the agent’s settings.

See How to use Slack integration in Vaiz for connection instructions.

Click Continue.

Step 6: Review your agent

Once created, your agent appears in the space’s agent library. Its card shows the agent’s status, scope, and last run time.

Open the agent to review its configuration, update its settings in Edit, or check its behavior in Test. Before using it for routine work, make sure its instructions, triggers, scope, and allowed actions match your workflow.

Step 7: Track the agent’s run history

Open your agent from the space’s agent library and select the Runs tab to view its run history.

For each run, you can see:

• The date and time.
• How it was triggered and, for manual runs, who started it.
• Its status, such as Suggestions ready or Failed.
• The number of actions, when available.

Use the run history to monitor the agent’s activity, find runs with suggestions ready for review, and identify failed runs that need attention.

To inspect run details across the whole space, approve proposed actions in Ask mode, and trace tool calls, see How to review and manage agent runs in Vaiz.

Example use cases

Use these examples as starting points. Adjust the instructions, scope, and allowed actions to fit your workflow.

Daily project briefing

Schedule an agent to review open tasks each morning and identify items that need attention.

Example instructions:

Review the project’s open tasks. List overdue items, tasks due in the next three days, and reported blockers. Include the assignee and a task reference for each item. Put urgent items first.

Weekly progress report

Run an agent at the end of each week to prepare a project update. Select a Slack channel if you want results posted there.

Example instructions:

Prepare a project update for the past seven days. Group it into Completed, In progress, and Needs attention. Include task references and highlight decisions the team needs to make. Keep it under 300 words.

Backlog preparation

Run an agent before a planning session to find unclear descriptions, missing acceptance criteria, and potentially overlapping requests.

Example instructions:

Review the backlog tasks. Flag missing goals, unclear scope, and missing acceptance criteria. Group requests that may overlap and explain why. Return a prioritized list of questions for the next refinement session.

Task description review

Run an agent manually to check whether a task is ready for work. Enable Write comments if you want feedback added to the task.

Example instructions:

Check whether the task has a clear goal, scope, and acceptance criteria. If information is missing, add a comment with specific questions for the author. Do not repeat questions already raised in existing comments.

Review handoff

Run an agent when a task is ready for review to help the reviewer understand what to check.

Example instructions:

Read the task description and available updates. Prepare a handoff summary covering the reported work, acceptance criteria to verify, and unresolved questions. Include references to relevant documents or deliverables.

Meeting follow-up

Run an agent after a meeting to turn notes into follow-up work. Enable Create tasks and use Ask mode to review proposed tasks before they are created.

Example instructions:

Read the provided meeting notes and propose a task for each action item. Include a clear title and a short description. Include an owner or deadline only when the notes specify one. Flag missing details for review and avoid duplicating existing tasks.

Release milestone planning and blocker linking

Schedule an agent or run it before a release cycle to group related tasks into a milestone, align start and due dates, and wire up blocker dependencies. Enable Manage milestones and Edit tasks, and use Ask mode so you can review the proposed milestone structure before it is applied.

Example instructions:

Review open tasks on the Core Platform and Web App boards that are not assigned to a milestone yet. Group related tasks for the upcoming release into a milestone with a clear title, subtitle, and target date range, and assign those tasks to the milestone. If a task depends on another task based on its description, comments, or linked pull requests, set the blocker relationship between them.

Following up with assignees on stalled or overdue tasks

Schedule an agent to check open tasks across a project, spot stalled or overdue work, and nudge the responsible assignees right inside the task thread without spamming them. Enable Write comments.

Example instructions:

Every weekday morning, check open tasks in this project that are not in Done, Closed, or Archived groups. Find tasks that are overdue or have had no updates or comments in the last 4 days. Before commenting, check the task’s recent comments—if you already left a reminder in the last 24 hours, skip the task. Otherwise, post a short, friendly comment mentioning the assignee with the task key, current board group, and due date (or how many days it is overdue), and ask what is needed to move it forward. In your summary, list every task you nudged and highlight overdue items first.

Writing and updating feature documentation from tasks and code

Trigger an agent when a task is updated (or run it at the end of a sprint) to create or update feature specifications and release docs from completed tasks, linked pull requests, branches, and commits. Enable Create & edit docs and Write comments.

Example instructions:

When a task on the Product Engineering board moves to Done, read the task description, comments, and its linked GitHub pull requests, branches, and commits. Find the corresponding feature page in the project document tree (or create a new document under Feature Specs if none exists). Update the document with what changed, configuration details, and user-facing behavior based on the task and merged code. Leave a short comment on the task with a link to the updated document.

Breaking down epics into subtasks and dependencies

Trigger an agent when it is assigned to a task or mentioned in a comment so your team can call it on demand to decompose a large feature into actionable subtasks. Enable Create tasks, Edit tasks, and Write comments, and use Ask mode to approve the proposed subtasks.

Example instructions:

When mentioned or assigned on a feature task, read its specification and linked documents. Propose a set of concrete subtasks covering backend, frontend, and QA verification. Give each subtask a clear title, acceptance criteria in the description, and an appropriate task type. Set blocker relationships between subtasks that must happen in sequence, and leave a comment summarizing the proposed breakdown.

Capturing architectural decisions and risks in Space Memory

Run an agent when comments are added or tasks are completed to keep your Space Memory up to date with durable technical decisions, conventions, and risks. Enable Write Space Memory.

Example instructions:

Review newly completed tasks and technical discussion threads in this project. When a thread settles an architectural trade-off, API contract, convention, or operational risk, search Space Memory first to avoid duplicates, then create a Memory entry of the matching kind (Decision, Convention, Context, Problem, or Risk) and link the related tasks.

For workflows built around predefined triggers and actions, see How to set up automations in Vaiz.