-
Notifications
You must be signed in to change notification settings - Fork 18
Workshop Agents
Agents are the durable AI participants of the Workshop. This page covers the Agents workspace: creating an agent, how definitions and conversations relate, revisions, enabling an agent to run, runtime settings, archiving, and how agents behave inside channels. For everyday conversation mechanics, see Workshop Collaboration Basics.
An agent has two halves with very different sharing rules:
- The definition is shared. An agent's handle, display name, purpose, instructions, and declared capabilities form one canonical record. Every workshop member can see and enable an active definition, but only the owner can change one; drafts stay visible to their owner alone. Definitions are versioned; each revision is immutable once saved.
- The conversation is private. When you enable an agent, you get your own direct lane with it, owned by you. Your conversations with the agent, and everything it remembers about you, belong to you alone. Another person enabling the same definition gets a separate lane and a separate memory; nothing crosses over.
So "the same agent" for two people means the same instructions, not the same experience. Each person's copy runs under an execution profile, accumulates its own memory, and can be archived independently.
Open Agents from the sidebar. The header shows how many agents are enabled and how many are active, the live connection state, and a Create agent button. Selecting an agent opens its detail view:
- Avatar, display name, @handle, and description.
- A lifecycle badge: active, draft, or archived.
- A button to start or open your conversation with it.
- An ownership note: whether you own this agent or someone else does. Either way, conversations and memory stay private per person.
- The active definition revision with its capability tags, the full revision history, the revision editor, and runtime settings, covered below.
Create agent opens a dialog with the fields that make up the first definition revision:
| Field | What it is |
|---|---|
| Handle | The agent's @name in mentions and autocomplete. Lowercase: a letter followed by letters, digits, or underscores, up to 32 characters. Handles share one namespace with people. |
| Display name | Human-readable name shown in the timeline and sidebar. |
| Avatar | The symbol shown next to the agent's messages. |
| Description | One-line summary shown in agent lists. |
| Purpose | What the agent is for; its charter. |
| Instructions | The standing directives that shape every response. |
| Declared capabilities | Checkboxes describing what the agent is built to do (see below). |
The dialog asks before discarding unsaved work when you close it. The creator becomes the agent's owner: revising, activating, and archiving the definition are owner actions.
Definitions are versioned, and revisions are immutable: saving never edits an existing revision, it creates a new one.
- The detail view shows the active revision, the one conversations actually use, and the list of all revisions.
- New revision opens an editor prefilled from the latest revision: Purpose, Instructions, and Declared capabilities.
- Saving a revision does not put it into use. A saved revision sits in the list until you press Activate on it. This is deliberate: you can draft, compare, and roll back by activating any revision, old or new.
If an agent is not behaving the way its newest revision says it should, check which revision is active first.
Each revision declares capabilities from a fixed set:
| Capability | Meaning |
|---|---|
| Text generation | Create ordinary text responses. |
| Tool activity | Expose bounded tool activity in the run inspector. |
| Workspace execution | Work within an already-authorized workspace. |
| Image input | Accept images when the selected runtime supports them. |
| Agent delegation | Delegate bounded tasks to other active agents in a shared channel. |
Capabilities are descriptive, not permissive. Declaring a capability grants nothing; it does not add tools, credentials, or data access. Execution profiles and workspace authorization, which operators control, decide what an agent can actually do. Capability tags exist so people browsing agents know what a definition is built for.
A definition on its own does nothing. To talk to an agent, you enable it, which activates the shared definition into your own private direct lane:
- Open the agent in the Agents workspace and choose Enable agent on an execution profile. The profile determines the backend and execution identity the agent runs under; an advanced disclosure lets you pick a specific profile when more than one is available.
- A direct message channel with the agent appears in your sidebar. It is owned by you and invisible to everyone else. Group channels are separate: an agent joins a group channel when the channel owner attaches it (see Agents in channels below); your personal enablement plays no part there.
Enablement is per person. Disabling or archiving your lane does not affect anyone else's use of the same definition.
The agent detail view includes inline runtime settings for your lane: backend, model, and response timeout, each with Apply and Reset to policy. Defaults come from the operator's runtime policy; you only need to touch these to override them, and Reset to policy returns to the operator-managed default. Workspace settings (active workspace and per-workspace overrides) live alongside them.
The same settings are reachable from the Settings view; links that point at an agent's settings land in the Agents workspace runtime section.
In group channels, agents are attached by the channel owner and sponsored by a person. The sponsor is the member whose enabled runtime backs the agent in that channel; the context pane's Agent attention section names the sponsor for each attached agent.
Each attached agent shows one of three attention states:
| State | Meaning |
|---|---|
| Available · not engaged | Attached and ready; not currently participating. Mention it to engage it. |
| Awake in this channel | Engaged; it is following the conversation and may respond without a fresh mention. |
| Unavailable until its sponsor re-enables this runtime | The sponsor's runtime for this agent is disabled, so the agent cannot run here until the sponsor re-enables it. |
Dismiss puts an engaged agent back to sleep without removing it from the channel; it stays attached and can be mentioned awake again. Detaching an agent entirely is a channel-management action for the owner (Manage channel agents).
Agents with the Agent delegation capability can hand bounded sub-tasks to other active agents in a shared channel. The server enforces hard limits on delegation, and every delegated run is visible in the channel like any other run.
Archive definition retires an agent. Archived agents keep their history: your direct lane with an archived agent becomes read-only rather than disappearing, and the definition remains in the archive dialog for review or restore. Archiving is an owner action, and like revisions it is reversible.
- Declared capabilities grant nothing. They describe intent; actual authority comes from execution profiles and workspace authorization.
- Conversations and memory are private per person, even when the definition is shared. Two people using the same agent share its instructions and nothing else.
- A revision must be activated after saving. Until you press Activate, conversations continue on the previously active revision.
- Workshop Collaboration Basics -- channels, threads, mentions, runs, and the composer.
- Multiple Backends -- the backends an execution profile can run an agent on.
- Memory -- how agent memory works under the hood.