The admin agent
Administer ContextWorks by conversation: draft briefs from a document, check health, compare two entities, publish a draft after approving its exact content. It runs as its own Agentforce agent or MCP server, separate from the runtime tools.
Set it up on Deploy → Agentforce, step 3a or Deploy → MCP, steps 2a–6a. You need the ContextWorks Admin or ContextWorks Domain Manager permission set, and every call runs as you — a domain manager gets their own domains, not the org.
How it works
Three rules govern what the agent may do:
- Reads need no approval. Health, the change log, the catalog, the configuration graph — aggregates and configuration, never record data.
- Drafts reach no one. The agent composes briefs, skills, and entity text freely, because a draft serves no one until a person activates it.
- Publishing previews first. Activating, adding to a domain, and deactivating each show a preview: exactly what will serve, or what the domain will carry. The agent’s instructions tell it to show you that preview and commit only with your yes.
- The commit must match the preview. Activating and adding to a domain return a digest. If the draft or the domain changes after the preview, ContextWorks refuses the commit, and the agent has to preview again. ContextWorks can’t check that you saw the preview.
What you can do
| You want to… | The agent uses |
|---|---|
| See what agents receive, per domain | Get Catalog |
| Check usage, quality, issues, trends, or writes | Get Health |
| See who changed what, and why | Get Change Log |
| Audit the configuration | Get Full Config |
| Draft a brief, a skill, or entity text | Save Draft |
| Publish a draft | Activate Draft |
| Change what a domain carries | Update Domain Contents |
| Pull something from service now | Deactivate |
| Run a packaged procedure | Get Skill · the skills |
Limitations
The agent cannot delete artifacts, change settings, grant access, or touch the write-access switches. Those stay in the Studio and Admin tabs.
Entity structure — fields, related lists, levels — is Studio-only too. The agent drafts an entity’s text: its description, instructions, and synonyms. It can’t create an entity, so the entity has to exist already.