Start here
Core concepts
A workspace is your company's home in Adapt: its agents, connections, documents, skills, and members. These are the terms used across the Adapt docs.
Workspace
Everything in Adapt belongs to one workspace. Agents, documents, skills, and enabled connections are shared company resources; conversations belong to the person who had them. Owners and admins set the workspace up in Settings, and everyone else uses what they set up.
Agent defaults (Settings → Defaults) are the starting settings for new agents: model, knowledge, sandbox, connections, and run limits. They're copied once when an agent is created; changing them later doesn't change existing agents. Company instructions are the exception: they apply to every agent at runtime, so editing them changes every agent at once. See Instructions and skills.
Agents
An agent is a named assistant with its own instructions, model, and capabilities. A new workspace starts with one company agent, named after the workspace. Any member can create more, and each agent can override the workspace defaults with its own model, sandbox, knowledge, and connections. See Agents.
Connections
A connection is a saved account with a provider: a model provider, a sandbox provider, a knowledge provider, an integration, or a surface. Owners and admins add connections; members choose from the enabled ones for agents they own.
- Models: what an agent thinks with. See Models.
- Sandboxes: isolated environments where an agent runs code and works with files. See Sandboxes.
- Knowledge: workspace documents, plus optional memory providers. See Knowledge.
- Integrations: the tools an agent uses in other apps, such as GitHub or Linear. See Integrations.
Sandboxes can also use secrets: environment values stored write-only in Settings → Secrets. See Secrets.
A connection marked Enabled is permitted for use. That's a permission, not a health check: it doesn't mean the credentials work or the provider is reachable.
Chats
A chat (conversation) is a thread between one person and one agent. Each message starts a run: the agent streams its answer and calls tools as it goes. Runs continue on the server if you close the browser, and stop when you choose Stop or the agent reaches its time or step limit.
Web chats are private to the person who started them: admins and owners can't read them, and history search only finds your own. An automation's runs are visible to anyone who can view that automation. See Chats and tool activity.
Tool activity
Every tool call an agent makes appears in the chat while it runs, then as finished or as failed with its error. A failed call didn't necessarily do nothing: if a request timed out or the run stopped partway through a call, the action may already have happened in the other system. Adapt never retries those calls automatically, and agents are told to say when an outcome is uncertain instead of claiming success. Check the other system before asking the agent to try again.
Owners and admins can review tool requests and setting changes in Settings → Activity log.
Automations
An automation runs an agent without someone typing a message: on a schedule, when you run it by hand, or when an event arrives from a surface. Scheduled and manual runs each get their own conversation, owned by the person who created the automation. See Automations.
Surfaces
A surface is a place outside Adapt where people talk to an agent. A message there becomes an ordinary run, and the answer goes back to the same thread.
- Slack: mentions, direct messages, and channel replies.
- GitHub: mentions in issue and pull request conversations.
- Email: an inbox the agent answers.
- iMessage and SMS: a phone number the agent texts from.
- Webhooks: events from any system that can send an HTTP request.
By default, surfaces only answer people who are members of the workspace. Unlike web chats, a surface thread can be opened read-only by any member who can see its agent.
Roles
Each member of a workspace has one role:
- Owner: created the workspace. Has full access, and the role can't be changed.
- Admin: manages settings, connections, documents, skills, people, and every agent.
- Member: uses agents, and creates and manages their own. Can view connections, documents, and skills, but not change them.
No role can read another person's web chats; an automation's runs are visible to whoever can view the automation. See Members and roles.