Skip to main content
Agents are persistent workers with an identity, purpose, instructions, model setting, capabilities, memory scope, workspace, and access boundary.

​
Agent types

​
Status

​
Identity and purpose

An agent’s name and title help people recognize it. Its purpose tells both people and other agents when work should be handed to it. Write purpose as a routing contract:

​
Configuration

The agent configuration panel covers:
  • identity, title, purpose, and visual profile
  • role and operating instructions
  • model selection
  • apps, skills, MCP servers, and other capabilities
  • team and delegation relationships
  • knowledge and memory behavior
  • channels and advanced runtime settings
Changes to protected system identities are intentionally restricted.

​
Delegation

Fleet reporting edges determine persistent delegation targets. A Team Agent can hand work to its direct workforce; a specialist can delegate further only when it has eligible direct reports and delegation is enabled. Delegation does not grant unlimited authorization. The target agent’s exact boundary-crossing tool calls still pass through automatic review and may pause for a human.

​
Tools and review

Agents can use connected apps, skills, MCP servers, workspace files, schedules, knowledge tools, browser and sandbox capabilities, and workflow tools when enabled. The capability existing on an agent does not mean every call executes automatically. SuperEmber reviews boundary actions against user intent and risk. See Automatic review and escalations.

​
Model selection

Each agent can follow the workspace default or use a supported override. Model availability, context, pricing, and defaults can change, so use the model picker as the current source of truth rather than relying on a fixed list in documentation.