What is an agent?
An AI coworker with its own instructions, tools, memory, and persistent workspace.
default agent agent access assistant team agent avatarOn this page
An agent has a place to workWhat an agent carries between tasksFrom a request to a resultYour agentsGive people accessDelegated workRecognize your agentsConfigure and use the agentAn agent has a place to work
An Opser agent is an AI coworker that can take actions to complete a task. It combines a language model with instructions, tools, memory, and a persistent filesystem. It can investigate information, write and run code, work with business records, create documents, and build apps when those capabilities are enabled.
The chat header shows the agent’s avatar — a friendly sky-blue character — and its name. The agent inspects available data and saves a customer portal plan as portal-plan.md in its persistent workspace.
The model handles reasoning and language. The surrounding Opser runtime supplies the workspace and the actions it can take. Switching models does not turn a different conversation into the same thing as a different agent.
What an agent carries between tasks
| Foundation | How it is used |
|---|---|
| Identity and instructions | Its name, behavior, role knowledge, and standing guidance |
| Business and personal context | Information about your organization and the people it serves |
| Memory | Facts the agent can save and retrieve for future work |
| Persistent files | Source files, scripts, notes, and results kept in that agent's workspace |
| Tools | Actions for reading, calculating, updating, building, and communicating |
| Skills | Instructions for carrying out a kind of work with those tools |
| Connections and access | The external accounts, apps, and records it is authorized to use |
Each agent has one ongoing chat, shared by everyone who can use it, and it never resets. When it gets long, the agent keeps a summary of earlier messages and looks up exact details in the chat when it needs them; its saved files and configured context stay as they are.
From a request to a result
An agent can break a task into steps, inspect inputs, choose tools, ask for a missing detail, run a script, and review the output. Some actions require approval. The chat shows what the agent writes and makes, with a short line saying what it is doing while it works; turn on its steps when you want to inspect each action. See what you see while the agent works.
For example, an agent can read several source files, calculate a comparison, save a reusable script, and produce a report. If the result belongs in an installed app, it can use that app's supported tools to create or update the record. Read Built-in tools and Code & scripts.
Your agents
You start with a default agent and can create additional agents for different kinds of work, such as a product manager, researcher, or blog writer. Each agent has its own workspace, configuration, and connections, and an owner, who controls its settings. On a specialist agent, the people with access can also teach it what to know and how to act (see Context & memory). Every plan lets you create as many agents as you need.
An agent does not gain access to another person's private files or an unconnected service just because you mention them. Ask it to identify the missing source or permission when it cannot complete a task.
Give people access
Any agent can be given to other people, including the default agent you start with. Its owner or an admin opens Access in the agent's settings and chooses Add people, or asks the agent in its chat, which asks for approval before giving or removing access. Access lists everyone who can use the agent, starting with its owner. An agent can serve a team around a process such as coordinating requests or preparing operational reports.
Everyone with access uses the agent together: one chat that everyone sees and writes in, where each message shows who wrote it, plus the agent's automations and their results, its files, its memory, and the accounts it is connected to. Connections belong to the agent, not to the person who made them. Each person keeps their own unread state.
When anyone besides its owner can use an agent, the faces of the people with access show at the top right of the chat, just before the search button, on the chat page and in the side chat. Select them to see who has access; its owner or an admin also sees Manage access there, which opens Access. When someone gets access, a line in the chat says so where it happened: they can see the whole chat, earlier messages too.
An agent never does more than its owner could, and never more than the person talking to it may do in Opser. Giving people access is not a shortcut around record permissions. Before you add someone, review what the agent's chat, files, and memory already hold and which accounts it is connected to: that person will be able to see and use them.
Open Settings → Agents to manage the agents available to you. Creating agents requires permission. Giving other people access to an agent is included on every plan, Free included.
Delegated work
For a longer task, the main agent can delegate a bounded part of the work to a subagent and bring its report back into the conversation. A subagent receives a task brief and selected tools; it is not another permanent agent you need to configure.
Delegation does not expand authority. Subagents cannot ask you questions or complete approval-gated actions themselves. The main conversation handles those decisions and brings the results together.
Recognize your agents
Every agent has a saved identity: a name and an avatar built from a shape and a color, with an expressive face. Your default agent starts as “Agent” with a randomly assigned avatar; when creating another agent, you can give it a role-specific identity. The General settings let you rename it and pick another look. Everyone who can use an agent sees the same identity.
Three agents, each with its own avatar: the Opser agent, used only by its owner, as a sky-blue rounded character; a Support agent four people use, as an apricot flower shape; and a Reports agent two people use, as a mint star shape.
The workspace sidebar lists the agents available to you above Apps, each with its avatar, name, and unread activity. Choose one to open its chat; an empty chat greets you with that agent's avatar, and chat headers show the avatar beside the name. Avatar motion follows your reduced-motion preference.
Configure and use the agent
The chat's header has a search button and a gear. The gear opens the agent's settings in a panel beside the chat, so you never leave it. It shows the agent, then its sections, each with its icon: General, Plugins, Approvals, Automations, Files, and Access. An open section shows a back arrow, its icon and its name, then a line saying what it is for; the arrow goes one level up. Drag the panel's edge to resize it and close it with ×; on a phone it opens as a sheet. The agent can change most of these settings itself when you ask in the chat. The controls available depend on your access.
Continue with your chat with an agent, context and memory, or persistent files.