# AI agents

Agents in Volobox are members of the workspace, not a chat window bolted onto it. Each person has a private assistant, and the workspace has shared team agents that people mention in chat, assign tasks to and put to work in workflows. An agent acts only within the access it is given, and its changes wait for a person's approval unless a standing rule or its authorization mode allows them.

![A chat with the personal assistant: the chat list on the left, a finished answer summarizing where the Fieldwork pilot stands, and an approval card proposing to create a Fieldwork task assigned to Leo Martinez, with Review & create, Create and Reject buttons.](https://volobox.io/docs/images/agent-chat.light.webp)

*A chat with the personal assistant in Agent mode: an answer about a project, and an approval card to create a task with Review & create, Create and Reject.*

## Personal assistant and team agents

| | Personal assistant | Shared team agent |
| --- | --- | --- |
| Belongs to | One person; nobody else sees it | The workspace |
| Works with | That person's access | The access of whoever mentions it, or its own memberships when it works on its own |
| Reached from | The person's own **Agent Chats** | `$mention` in team chats, task assignment, workflows |
| Set up by | Created for every user | An administrator: handle, instructions, model, skills, members, manager |

In a private chat, a person's assistant can do nearly everything Volobox offers through its API, such as setting up agents and workflows or changing many tasks at once. Each change is an approval card.

> **Note:** No auto-approval rule covers the changes a personal assistant makes this way.

**Volly** (`$volly`) is a built-in shared agent everyone can mention. The **Agents** directory shows which agents exist, what each may do, which workflows it runs and what it has done lately.

![The Team page on its Agents tab: a search box, filters for interactive and autonomous agents and for roles, and six shared agents as cards (Content Writer, Project Guide, Release Manager, Software Developer, Studio Coordinator, Support Triage), each with its handle, the caller’s role, and Open to everyone, Interactive and Autonomous badges.](https://volobox.io/docs/images/agent-directory.light.webp)

*The Agents tab of the Team page: the shared agents as cards with their role and badges, and filters for interactive and autonomous agents and for roles.*

## Ask, Plan and Agent modes

- **Ask** answers and reads, and makes no changes to the workspace.
- **Plan** explores, asks questions, and ends with a plan you review. **Approve and run** then carries it out in Agent mode.
- **Agent** proposes changes, each on an **approval card**. Nothing runs until you approve it, unless a standing rule you set covers it.

When the change is a new meeting, task or request you asked for, its card opens the real create form, filled in with the agent's draft (**Review & create**). You change what you like and create it there, and the agent is told what was created and which fields you changed.

## Approvals, standing rules and safety limits

A shared agent acts in one of three authorization modes:

- **User delegated** acts only within the access of the person using it.
- **Manager escalation** sends a write that person cannot do to the agent's manager for approval.
- **Self-authorized** lets an agent working on its own act within its own memberships, without cards, inside its safety limits.

A **standing rule** pre-approves matching actions so you are not asked every time. A rule can be narrowed to a project, person or space, and can expire. It can also carry a **budget** per hour, per day or in total; once the budget is spent, the next call comes back as a normal approval card.

Separately, an agent's **safety limits** cap how many writes it may make per run or per hour when it works on its own.

> **Note:** A reply to a customer is never auto-approved. A person always sends it.

![The Software Developer agent’s profile: its role, Disabled in chat, Self-authorized for unattended work with Maya Chen as manager, its write ceiling, a Skills card listing the built-in skills and the Meridian release process, QA checklist and Fieldwork repository skills, and the start of its charter.](https://volobox.io/docs/images/agent-settings.light.webp)

*A shared agent's profile: how it works in chat and when unattended, its manager, its limit on writes, its skills and the start of its charter.*

## Skills, commands and packs

**Skills** bundle tools and instructions for a domain, and an agent loads them when it needs them. Builder skills (**Agent Builder**, **Workflow Builder**, **Insights Builder**, **Workspace Admin**) let your assistant set up agents, workflows, dashboards and workspace settings from a private chat.

Skills can be imported and exported as `SKILL.md` files or zip bundles. An imported skill grants no permissions by itself. **Slash commands** are reusable prompt shortcuts such as `/summarize`.

![Organization Settings on the AI tab, Skills: the Skill library with Organization and System tabs, Import and Create skill buttons, and Meridian Works skills as cards (Brand voice, Client communications, Fieldwork repository, Meridian release process, QA checklist, Research synthesis), each with who can use it, its description and its tools.](https://volobox.io/docs/images/skills-library.light.webp)

*The Skill library in Organization Settings: the organization's skills as cards, with Import and Create skill.*

**Library packs** bundle agents, skills, commands and connections. Volobox ships packs such as Software Delivery Team, Release Manager, Campaign Planner, People Ops Partner and Chief of Staff. You can also publish your own, or pull them from a registry you configure.

Packs are never installed automatically, and installed items are ordinary resources you can edit. Pack agents prepare and propose; people decide and publish.

![Organization Settings on the AI tab, Library: the Meridian Works packs registry with its index address, and pack cards such as Software Delivery Team, Release Manager (installed, one edited item) and Delivery Coach, each with its version, a Verified badge, its description, what it produces and example prompts.](https://volobox.io/docs/images/library-packs.light.webp)

*The Library in Organization Settings: a connected pack registry, and pack cards with their version, badges, description, example prompts and Install.*

To write your own, follow [Create a skill](/docs/create-a-skill) or [Host a pack registry](/docs/host-pack-registry). The file layouts are in [Skill and pack formats](/docs/skill-and-pack-formats).

## Memory, guidance and learned notes

Your assistant can **remember** durable facts about you, when you ask (or use `/remember`) or when it notices a preference. Each memory shows in the chat with **Undo** and belongs only to you. A chat can be marked **Don't remember this chat**.

For a project, **agent guidance** is what you instruct agents to do there, and **learned notes** are what the team agreed is true ("We release on Thursdays"). An agent proposes a note, a project Admin approves it, and every agent on the project then knows it. When the two disagree, guidance wins.

## Sources, artifacts and charts

A chat with an agent can keep **sources**: uploaded files, web pages and Volobox items. A Volobox item is never copied: the agent reads it with the access of whoever is asking, each time.

The agent can publish **artifacts** (HTML, Markdown, CSV or JSON documents) in any mode, keep versions as you ask for changes, and save them to Drive or as a Docs page.

When an answer has numbers, the agent can draw a **chart card**. A card is a snapshot of what the agent read. **View data** shows the rows and where they came from, and **Build an insight like this** turns it into a chart that refreshes. See [Dashboards and alerts](/docs/dashboards-and-alerts).

## Voice chat

In **voice chat** one person talks out loud with their assistant or a team agent in their own chat with it. What both say is written into the chat, marked **Said aloud**, and approval cards still apply.

Voice chat is not a call. Agents in calls are covered in [Chat and calls](/docs/chat-and-calls).

## Channel agents

A **channel agent** is a workflow that watches one channel and acts on new messages without being mentioned, for example turning bug reports into tasks. It runs as any workflow does. See [Workflows](/docs/workflows).

## Knowledge search

Agents search the workspace's knowledge: Docs, task and request text, meeting notes, pull requests, and tickets on the desks the searcher staffs.

Volobox Help is a separate search, so an agent can also explain how Volobox itself works.

## Related

- [Workflows](/docs/workflows)
- [Coding](/docs/coding)
- [Chat and calls](/docs/chat-and-calls)
- [Security and administration](/docs/security-and-administration)
- [Create a skill](/docs/create-a-skill)
- [AI Agents](/product/ai-agents)
