Agents are AI participants inside a Craken workspace. They can talk with people, read connected context, and use workspace tools to produce durable results such as wiki edits, summaries, file analysis, and follow-up messages.
Every new workspace includes Orca. Admins can add other server-managed agents, and users can also connect external CLI agents through the CLI.
How To Call An Agent
Use one of these paths:
- Open Direct messages to Orca for private work.
- Mention
@Orcaor another agent in a channel where that agent is a member. - Include links to the relevant pages, files, or channels when the request depends on existing context.
Agents do not usually interrupt ordinary human conversation unless they are mentioned or the thread is clearly addressed to them. If an agent does not answer in a channel, check the mention and channel membership.
What Agents Can Do
Agents can perform workspace actions through tools:
- Read recent channel and DM context, search older messages, and answer in the current conversation.
- Find, read, write, revise, rename, restore, and verify wiki pages when the right wiki workflow is loaded.
- Inspect Shared files, read text and Markdown files, extract text from supported PDFs and office/data files, summarize evidence, rename shared files, and promote visible attachments into Shared files when asked.
- Connect decisions across channels, wiki pages, shared files, and messages.
- Search and inspect academic literature when the research skills are available.
- Create visible Jobs for long-running work and update the plan as progress is made.
- Remember stable workspace preferences and useful workflows over time.
For a capability-focused reference, see Orca Capabilities.
Give Useful Context
Good requests include the goal, the source material, and the desired output. Link the source material directly:
Based on last week's discussion in #research, update the Risks section of [[Launch Plan]].
Use [[/briefs/q2-risks.pdf]] as evidence and return the result as a short table.
Jobs And Progress
Longer work becomes a Job. The conversation shows a progress row and a plan/checklist. Admins can also inspect Jobs from Manage Workspace. If you want to change direction, send another message in the same conversation. To stop explicitly, use the interrupt control.
Token Limits
Agent usage is shared by the workspace and appears in the navigation footer. Simple questions use less of the quota; long, multi-step work uses more. If the workspace is near a limit, break work into smaller requests or wait for the next daily or monthly reset.
Learning
Agents can learn stable preferences and successful workflows from the workspace. Admins can inspect Dreaming outcomes in Manage Workspace.