Agents
An agent is an AI identity that acts inside one organization. CortexHub keeps the identity, runtime, ownership, capabilities, and governance boundary explicit so admins can see what is acting and why.
The four agent types
| Type | Who owns it | Where it runs | Typical use |
|---|---|---|---|
| Hosted Org Agent | Organization | CortexHub | Shared work for the organization, including a named Slack bot |
| SDK Org Agent | Organization | Your runtime | A partner or internal application acting as the organization |
| Co-worker | One member | CortexHub | Private assistant for that member's chat, files, routines, and meetings |
| External client | Member or end user | Claude, Cursor, ChatGPT, or another runtime | Bring your existing AI client through the governed MCP gateway |
Org Agents have no member owner. Their toolkit access comes from an agent group, and their activity is attributed to the organization. Co-workers use a member's connections and remain private to that member. External clients connect through MCP OAuth or an org gateway key, depending on the integration.
When to use each type
- Use a Hosted Org Agent when the organization should run a shared assistant and members should be able to mention it in Slack.
- Use an SDK Org Agent when your application owns the runtime but the identity, policies, and audit trail should belong to the organization.
- Use a Co-worker for member-private work that uses that member's files, Brain, meetings, routines, and connected tools.
- Use an External client when a person wants to keep using Claude, Cursor, ChatGPT, or another MCP-capable app.
All four types use the same governance pipeline before a tool runs: entitlement, safety profile, policy decision, obligations, approval, and audit.
How Org Agents are governed
- An admin creates or registers the Org Agent.
- The agent is assigned to an agent group.
- The agent group receives toolkit entitlements and tool policies.
- The runtime discovers and invokes tools through CortexHub.
- CortexHub permits, parks for approval, or denies each action and records the result.
An agent group is separate from a member group. This lets an organization give shared agents a controlled tool set without changing what employees can connect for their own work.
Slack identity
A hosted Org Agent can be installed as a named Slack bot. The admin connects Slack once for the organization, enables named bots, then selects Add to Slack from the agent's Slack settings. The bot can be added to channels and mentioned by name. Installing the bot does not grant tools; the agent group and policy still decide what it may do.
Slack approval channels are a separate concept. They deliver operator decisions for parked actions. The named agent bot is the surface where members ask the agent to work.
Identity and credentials
Hosted Org Agents run on CortexHub and do not need a runtime API key. SDK Org Agents run in your environment and use an org API key bound to the agent. External MCP clients use member OAuth or an org gateway key with end-user attestation. Co-workers use the owning member's connections.
Agents never receive raw vendor credentials from connected toolkits. CortexHub resolves the connection at invocation time, applies policy, and records the governed outcome.
Lifecycle
| State | Meaning |
|---|---|
| Active | The agent can connect and invoke tools subject to policy. |
| Blocked | An admin has temporarily stopped access; history remains available. |
| Archived | An Org Agent is stopped while its history is preserved. |
Restore an archived agent only after reviewing its agent group, capabilities, Slack installation, and policies. Purge is owner-only and permanently removes the agent's dependent history.