Assistants
Save useful AI behavior once, equip it for a clear job, and let the intended people use it in their own work.
An assistant is an AI helper whose saved instructions and tools can be reused. Mobius calls this resource an agent in the app, API, and CLI, so those exact labels and identifiers remain in this page.
Northwind's Scout assistant can turn supplied service notes into an operations brief. Maya improves Scout's instructions once, then selected colleagues can use the same Scout in separate conversations. The useful behavior is shared; their messages are not combined.
What you save on an assistant
Open an assistant's Configure tab to see the choices that shape its work:
| Setting | What it changes |
|---|---|
| Agent instructions | The job, expected result, and rules for missing information |
| Model | The AI that handles the work |
| Tools | The actions the assistant may call |
| Accounts | The named connected accounts available to those actions |
| Skills | Extra instructions it may load for a particular task |
| Memory | Facts deliberately saved for later conversations |
| Who this agent is for | The people who may find and use it |
The organization can also hold shared background under Build > Context. Every member may read that company context, only administrators may change it, and it is included in later assistant work. Keep it to facts that every assistant in the organization should receive.
An assistant does not get an external account merely because that account is connected. Grant the assistant access to the exact account it needs. The same grant can support eligible conversations and routines, while the account's acting identity and allowed result audience still apply. See connect accounts and set access.
Ask another of your assistants
When enabled for your organization, an assistant you are talking with privately can ask any other assistant you can use: your individual assistants and the team assistants available to you. For example, tell Inbox Assistant, “Ask Research Assistant for the vendor shortlist, then draft my reply.” Research Assistant answers as it would in a private chat with you, using its own instructions, your memory, and the accounts you have allowed it to use.
The Asked Research Assistant step opens a private, read-only record of the exchange. Later questions from Inbox Assistant continue that same exchange. These records stay out of your chat list. The assistants ask each other; they do not read each other's conversations.
The assistant does not wait for the reply. It tells you it asked, and you can keep working with it. The step shows Research Assistant is working on it with the time so far. When you scroll away from the step, a line above the message box shows the same status, with View to jump back. The reply arrives later in the same conversation, as Reply from Research Assistant, and your assistant picks it up from there. When the next step needs the answer first, the assistant waits for it, for up to 15 minutes. You can send another message while it waits. It then stops waiting, and the reply arrives later in the same conversation. A routine, and an assistant that is answering another assistant, always wait for the reply. A request can involve up to five questions across your assistants, with a total limit of 30 questions per hour. Their work counts toward your own usage limits.
While one assistant works on another's question, the asking assistant can pass on a correction. The answering assistant can also send a question back. It shows in your chat as Question from Research Assistant, and the step reads Research Assistant is waiting for an answer until your assistant replies. Your assistant may ask you to answer it.
Each question returns one answer. If an assistant consults a third assistant and answers before that reply arrives, the later information stays in their exchange. It does not send you a revised answer.
This works only where the assistant is working privately for you: your own chat, DM, or individual routine. Shared chats and team routines can't ask, and other people's individual assistants are never asked. Only you can open the exchange, even when the other assistant is a team assistant. Administrators can turn Assistants can ask each other off in organization settings. Existing exchanges remain readable.
Create a focused assistant
You need permission to build resources. Making an assistant available to a group or the whole organization also requires organization administration.
- In the app, open Build > Agents.
- Click New agent.
- Fill in Agent name, Agent description, and Agent instructions.
- Leave Model on the available default for the first test.
- Under Who this agent is for, leave Just me selected.
- Click Create agent.
The app creates a private assistant by default. The employee API does the same
when visibility is omitted, but API clients should set that field explicitly.
A service principal that is preparing an assistant for other people must name
the intended organization or restricted audience rather than relying on a
default.
For a first version of Scout, these instructions are enough:
Prepare an operations brief from facts supplied in the request. Do not invent
missing owners, causes, or dates. Include Schedule and Risks. If a section has
nothing to report, say so.The instructions describe a job you can check. Scout has no reason to contact another service yet.
Use before you equip
On the assistant's page, click Chat and give it facts you already know. Check the reply for unsupported claims and missing details. Open a fresh conversation after an instruction change so the same input gives you a clean comparison.
Add a tool or account only when the test shows why it is needed. Tools describe what the assistant may try, while account grants decide which connected identity it may use. Mobius revalidates the exact connection, live grant, account binding, and container audience before every external action. It does not automatically ask a person to confirm an action that changes data. Add a deliberate interaction to the instructions, and tell the assistant to wait for the answer, when the task should stop for human review. That review follows the instructions instead of acting as an automatic gate on every write.
Share use without sharing the conversation
An assistant's audience controls who may reach it:
- Just me keeps it with one person.
- Only listed people gives a named group the same saved behavior.
- Everyone in the org makes it available across the organization.
Audience is one part of access. A listed person still needs the organization permission to use assistants. Reading configuration, editing the definition, managing the audience, controlling an account, and receiving an external result are separate decisions. Who can use an assistant shows how they fit together.
When two people use Scout, each conversation becomes a separate session. A routine occurrence also gets its own session. The routine's saved setup names Scout, but it does not continue in the private chat where someone first suggested the routine.
Improve the assistant people already use
Open Build > Agents, choose the assistant, select Configure, and click Edit. Change the instructions, model, tools, skills, or other settings, then click Save.
The assistant keeps the same ID. Earlier conversations and completed turns do not change. Future turns read the current saved definition, including the next turn in an existing conversation. A conversation with an explicit model override keeps that model even if you change the assistant's model default.
Edit the existing assistant when you are improving the same job for the same access boundary. Create a separate one when the new job needs different people, accounts, or private knowledge.
Put information in the right place
| Information | Where it belongs |
|---|---|
| Messages and facts for this conversation | The current session |
| A fact Scout should use in later work | Deliberately shared memory |
| A procedure several assistants should load | A skill |
| Structured working data | A table or file, or the system that owns it |
Mobius does not learn a new team rule from everyone's conversations automatically. Someone has to promote the useful change into instructions, shared memory, a skill, or a maintained data source.
Delete an assistant
Open Agent actions on its page and choose Delete agent when the company no longer needs that assistant. It stops taking new work; work already in progress is not cancelled. Check its routines, connected accounts, and records first so you know what will lose its active dependency.
A deleted assistant keeps its chats, memory, and instructions. An Owner or Admin can bring it back from Settings, Deleted items, with no time limit. It comes back on, with the same ID. Organization accounts, skills, the people it was shared with, and its managers come back when each is still valid. An individual assistant whose owner left the organization, for example one deleted or kept for the team at offboarding, cannot be restored, and the list says so. A person's own account comes back only when that person restores it; otherwise they share it again. Old API keys and CLI tokens never work again. Routines that the delete paused stay paused until someone resumes them. Chat apps it was set up to answer in, such as Slack or Teams, work again right away. After a restore, a notice on the assistant's page lists anything that did not come back and why.
Exact API fields live in the Agent configuration API guide. Terminal commands are in the CLI command reference. For a focused editing method, read assistant instructions and settings.