Complete provisioning workflow for Einstein Agent Users and permission sets. Validated against ORM1, ORM2, AutomotiveSupport, and SalesforceProductAssistant agents.
---
## License Requirement
PID_DigitalAgent (typically included with Agentforce licenses)
| **Data Cloud permset/PSL** (one of `GenieDataPlatformStarterPsl` PSL, `GenieUserEnhancedSecurity` PS, or `DataCloudUser` PS) | **Required when agent has `knowledge:` block** | Not needed |
| **`access.default_agent_user`** | Required | Omit unless a capability explicitly requires it |
> **Why the Data Cloud permset name varies:** which permset/PSL grants Data Cloud access depends on org shape (scratch / Dev Edition / Trailhead trial / sandbox / production) and platform release. Three names are seen in the wild:
> - `GenieDataPlatformStarterPsl` — a Permission Set License (assigned via `PermissionSetLicenseAssign`, not `PermissionSetAssignment`).
**How to check agent type**: Look at the `agent_type` field in the `config:` block of your `.agent` file, or query: `sf data query --json --query "SELECT DeveloperName, Type FROM BotDefinition WHERE DeveloperName = 'AgentName'" -o TARGET_ORG`
---
## CLI Fast Track: Complete Workflow
For CLI-first workflow (tested: ~8 minutes total):
```bash
# Step 1: Query existing Einstein Agent Users (30 seconds)
sf data query --json \
--query "SELECT Id, Username, IsActive FROM User WHERE Profile.Name = 'Einstein Agent User' AND IsActive = true" \
-o TARGET_ORG
# Step 2: Create Einstein Agent User (2 minutes)
# Get Profile ID (read result.records[0].Id from JSON response)
sf data query --json \
--query "SELECT Id FROM Profile WHERE Name = 'Einstein Agent User'" \
-o TARGET_ORG
# For Production/Sandbox (non-scratch org):
# Use the ProfileId from the query above
sf data create record --json --sobject User --values \
"Username=<agent_name>_user@<orgId>.ext \
LastName=<AgentName> \
Email=admin@example.com \
Alias=<alias> \
TimeZoneSidKey=America/Los_Angeles \
LocaleSidKey=en_US \
EmailEncodingKey=UTF-8 \
ProfileId=<PROFILE_ID> \
LanguageLocaleKey=en_US" \
-o TARGET_ORG
# For Scratch Orgs (use user definition file):
# sf org create user --definition-file config/einstein-agent-user.json -o TARGET_ORG
# Step 4: Deploy Custom Permission Set (3 minutes)
# (Create the .permissionset-meta.xml file first - see Section 3.2 template)
sf project deploy start --json \
--metadata PermissionSet:<AgentName>_Access \
-o TARGET_ORG
# Assign custom PS
sf org assign permset --json \
--name <AgentName>_Access \
--on-behalf-of <agent_name>_user@<orgId>.ext \
-o TARGET_ORG
# Step 5: Verify All Permissions (1 minute)
sf data query --json \
--query "SELECT PermissionSet.Name, PermissionSet.Label FROM PermissionSetAssignment WHERE Assignee.Username = '<agent_name>_user@<orgId>.ext' ORDER BY PermissionSet.Name" \
## Step 3b: Assign Data Cloud Access (Knowledge-Grounded Service Agents Only)
Run this ONLY when the agent has a top-level `knowledge:` block (i.e., it grounds answers on an Agentforce Data Library). Without Data Cloud access, the agent user cannot read the ADL's data space and `AnswerQuestionsWithKnowledge` returns empty `knowledgeSummary` — the anti-hallucination guard then refuses every utterance.
Skip this step for non-grounded agents.
### 3b.1 — Discover which Data Cloud permset/PSL exists in this org
Three names are seen in the wild; which one applies depends on org shape and platform release. Run both queries:
```bash
# (a) Look for the PSL form first (codey-cko2's preferred path):
sf data query --json \
--query "SELECT DeveloperName FROM PermissionSetLicense WHERE DeveloperName = 'GenieDataPlatformStarterPsl' LIMIT 1" \
-o TARGET_ORG
# (b) Look for the PS form(s) — the names that test-agent17 found working:
sf data query --json \
--query "SELECT Name, Label FROM PermissionSet WHERE Name IN ('GenieUserEnhancedSecurity', 'DataCloudUser', 'DataCloudArchitect')" \
-o TARGET_ORG
```
Pick **one** to assign, in this priority order:
1.`GenieDataPlatformStarterPsl` (PSL) — if (a) returned a record. Use the PSL flow.
2.`GenieUserEnhancedSecurity` (PS, label "Data Cloud User") — if (b) returned this name.
3.`DataCloudUser` (PS) — if (b) returned this name.
4.`DataCloudArchitect` (PS) — last resort; this is typically an admin permset (over-privileged for an agent user) but if nothing else exists in the org, it works.
If none of the four exist: Data Cloud is likely not provisioned. Run the Step 0 preflight from [Data Library Reference](data-library-reference.md) to confirm; if DC is missing, surface that to the user and skip ADL grounding for this run.
### 3b.2 — Assign
For a **PSL**, use `sf org assign permsetlicense`:
```bash
sf org assign permsetlicense --json \
--name GenieDataPlatformStarterPsl \
--on-behalf-of <agent_name>_user@<orgId>.ext \
-o TARGET_ORG
```
For a **PS**, use `sf org assign permset`:
```bash
sf org assign permset --json \
--name <PS_NAME_FROM_DISCOVERY> \
--on-behalf-of <agent_name>_user@<orgId>.ext \
-o TARGET_ORG
```
Both are idempotent — re-running on an already-assigned user returns success without effect.
### 3b.3 — Verify what landed (don't trust the apparent-success response)
Apex-driven assignments can silently roll back inside transactions, leaving the user with no permset while the API reports success. Read back from the assignment SObjects directly:
```bash
# Permsets (PS):
sf data query --json \
--query "SELECT PermissionSet.Name FROM PermissionSetAssignment WHERE Assignee.Username = '<agent_name>_user@<orgId>.ext'" \
-o TARGET_ORG
# Permission set licenses (PSL — different SObject):
sf data query --json \
--query "SELECT PermissionSetLicense.DeveloperName FROM PermissionSetLicenseAssign WHERE Assignee.Username = '<agent_name>_user@<orgId>.ext'" \
-o TARGET_ORG
```
The Data Cloud name you assigned in 3b.2 must appear in one of the two result sets. If it does not, the assignment failed silently — surface the failure and try the next-priority name from 3b.1.
### 3b.4 — Verify the assignment stuck (pinned to resolved IDs)
Run this in the main assignment flow, immediately after 3b.3 and before any code generation or preview. 3b.3 checks that a Data Cloud permset name appears for the agent user's `Username`. That is necessary but not sufficient: assignment writes can silently roll back inside an Apex transaction while the API reports success, and the "N rows for N users" shape of an unpinned query hides the specific miss. This step pins verification to the exact agent-user Id so a missing agent-user row is a HARD failure — not a row-count that happens to look plausible.
**The agent user is the one that must hold the assigned permset.** 3b.2 assigns the Data Cloud permset to the Einstein Agent User, which runs the retriever at runtime; if it is missing, grounding surfaces as empty `knowledgeSummary` and anti-hallucination refusals at preview/runtime. The running user (whoever runs `sf`, configuring the ADL) also needs Data Cloud access, but that access commonly comes from its profile (e.g. a System Administrator) rather than one of the discovered permsets — so the running user is a **soft** check here, not a hard gate.
**Resolve the agent-user Id first** (this is the pin):
```bash
sf data query --json \
--query "SELECT Id, Username FROM User WHERE Username = '<agent_name>_user@<orgId>.ext'" \
-o TARGET_ORG
# Read result.records[0].Id — this is the agent user's Id.
```
**Re-query the assignment SObject pinned to that Id.** Substitute the resolved agent-user Id and the exact PS/PSL name assigned in 3b.2:
```bash
# PS form (if 3b.2 assigned a PermissionSet):
sf data query --json \
--query "SELECT AssigneeId, Assignee.Username, PermissionSet.Name FROM PermissionSetAssignment WHERE AssigneeId = '<AGENT_USER_ID>' AND PermissionSet.Name = '<PS_NAME_FROM_3b.2>'" \
-o TARGET_ORG
# PSL form (if 3b.2 assigned a PermissionSetLicense):
sf data query --json \
--query "SELECT AssigneeId, Assignee.Username, PermissionSetLicense.DeveloperName FROM PermissionSetLicenseAssign WHERE AssigneeId = '<AGENT_USER_ID>' AND PermissionSetLicense.DeveloperName = '<PSL_NAME_FROM_3b.2>'" \
-o TARGET_ORG
```
**Assertion — read the returned `AssigneeId`, not the row count.** `<AGENT_USER_ID>` must appear in the `AssigneeId` column of the appropriate result. An empty result is a FAILURE — do not proceed on "a row came back" reasoning; the specific agent-user Id must be present.
**On failure:**
1. Re-run 3b.2 for the agent user, then re-run this 3b.4 verification once.
2. If the agent user is still missing after one retry, STOP. Surface which PS/PSL name did not stick. Do NOT proceed to code generation, deploy, publish, or preview — grounding will fail silently at runtime and the failure is easier to diagnose here than after a broken ship. If nothing in 3b.1's priority list assigns cleanly, treat this as a Data Cloud provisioning problem and re-run the Step 0 preflight from [Data Library Reference](data-library-reference.md).
When the agent-user Id appears for the intended PS/PSL, Step 3b's assignment gate is satisfied and callers upstream (SKILL.md, spec/orchestrator) can treat "Step 3b passed" as an authoritative gate on Data Cloud grounding access — no need to re-run inline SOQL from the caller. If grounded queries still return empty at runtime, apply the Data Space scope fallback in 3b.5.
### 3b.5 — Data Space scope (UI-only manual fallback if grounded queries still fail)
Some org shapes require a **separate** Data Space scope grant on the assigned permset, in addition to the assignment itself. There is currently no API for this — it must be done in Setup UI.
Do this only if 3b.1–3b.4 succeeded but the agent still returns empty `knowledgeSummary` at runtime:
> Setup → Permission Sets → click the assigned Data Cloud permset → "Data Cloud Data Space Management" under the Apps section → Edit → add the ADL's data space (typically `default`) to the **Enabled Data Spaces** list → Save.
>
> The data-space ID can be found via `sf data query --json -q "SELECT Id, DeveloperName FROM DataSpace"`.
After granting the scope, retest with a grounded utterance — the agent should now return populated `knowledgeSummary`.
**Note**: `sf org create user` only works in scratch orgs. For production/sandbox, use `sf data create record`. Attempting `sf org create user` in a non-scratch org fails with an authorization error.
4. Verify creation:
```bash
sf data query --json --query "SELECT Id, Username, IsActive FROM User WHERE Username = '{agent_name}_agent@{orgId}.ext'" -o TARGET_ORG
```
**Username format**: `{agent_name}_agent@{orgId}.ext` (production) or `{agent_name}.{suffix}@{orgfarm}.salesforce.com` (dev/scratch). Always query the target org to confirm the exact format.
---
### Step 2: Assign System Permission Set (`AgentforceServiceAgentUser`)
Critical: Must be assigned BEFORE publishing the agent. Without it, publish fails with "Internal Error".
sf data query --json --query "SELECT Id, PermissionSet.Name FROM PermissionSetAssignment WHERE Assignee.Username = '{agent_name}_agent@{orgId}.ext' AND PermissionSet.Name = 'AgentforceServiceAgentUser'" -o TARGET_ORG
```
---
### Step 3: Create Custom Permission Set for Apex Classes
The custom PS grants the agent user permission to execute your Apex invocable actions.
<!-- Add one entry per Apex class the agent calls -->
<classAccesses>
<apexClass>YourApexClassName</apexClass>
<enabled>true</enabled>
</classAccesses>
<!-- Repeat for ALL Apex classes referenced via apex:// in agent script -->
</PermissionSet>
```
Key rule: Include EVERY Apex class referenced via `apex://` in your agent script. Missing even one causes "invocable action does not exist" at runtime.
Deploy the permission set:
```bash
sf project deploy start --json --source-dir force-app/main/default/permissionsets/{AgentName}_Access.permissionset-meta.xml -o TARGET_ORG
```
---
### Step 4: Assign Custom Permission Set to Agent User
sf data query --json --query "SELECT PermissionSet.Name FROM PermissionSetAssignment WHERE Assignee.Username = '{agent_name}_agent@{orgId}.ext'" -o TARGET_ORG
**Validated workflow pattern**: Deploy as unpublished metadata, test with preview, then publish only when tests pass. This avoids version management overhead during iteration.
This deploys the agent as **unpublished metadata** — you can edit freely without version management.
#### 6.2: Test with Preview (Before Publishing)
```bash
sf agent preview start --json \
--api-name <AgentName> \
-o TARGET_ORG
```
What to test:
1. All subagents trigger correctly
2. All Apex actions execute without "Insufficient Privileges" errors
3. Agent responds with expected data
4. No compilation errors
If testing reveals problems, edit your agent script or Apex classes, redeploy, and test again — no publish required.
**⚠️ `WITH USER_MODE` Object Permissions:** Apex using `WITH USER_MODE` requires the Einstein Agent User to have read access on queried objects. Class-level access alone is not enough. Missing object permissions fail silently — 0 rows, no error. If live preview returns empty but simulated works, check Setup > Profiles > Einstein Agent User > Object Permissions. Fix by adding `<objectPermissions>` to your custom PS:
- Auto-generated `NextGen_ORM1_Permissions` only included 3 classes (missing `ShipmentTracker`)
- Runtime error: "invocable action track_delivery does not exist"
- Fix: Created custom `ORM1_Access` with all 4 classes — no errors
Best practice: Always create your own custom `{AgentName}_Access` PS with explicit `<classAccesses>` for every Apex class. Ignore the auto-generated PS.
---
## End-to-End Verification Checklist
Run this combined query to verify all setup steps for a Service Agent:
```bash
# 1. Einstein Agent User exists and is active
sf data query --json --query "SELECT Id, Username, IsActive, Profile.Name FROM User WHERE Username = '{agent_name}_agent@{orgId}.ext'" -o TARGET_ORG
# 2. System PS assigned
sf data query --json --query "SELECT PermissionSet.Name FROM PermissionSetAssignment WHERE Assignee.Username = '{agent_name}_agent@{orgId}.ext' AND PermissionSet.Name = 'AgentforceServiceAgentUser'" -o TARGET_ORG
# 3. Custom PS assigned
sf data query --json --query "SELECT PermissionSet.Name FROM PermissionSetAssignment WHERE Assignee.Username = '{agent_name}_agent@{orgId}.ext' AND PermissionSet.Name = '{AgentName}_Access'" -o TARGET_ORG
# 4. All permission sets for user (combined view)
sf data query --json --query "SELECT PermissionSet.Name, PermissionSet.Label FROM PermissionSetAssignment WHERE Assignee.Username = '{agent_name}_agent@{orgId}.ext'" -o TARGET_ORG
- **Prevention:** Step 6 splits publish and activate into separate steps with verification
- **Result:** Agent is both published AND activated
---
## Troubleshooting
| Error | Cause | Fix |
|-------|-------|-----|
| "Internal Error" on publish | `AgentforceServiceAgentUser` PS not assigned to Einstein Agent User | Assign system PS (Step 2), wait 2-3 min, retry publish |
| "Insufficient Privileges" at runtime | Custom PS missing or incomplete `<classAccesses>` | Verify custom PS includes ALL Apex classes, redeploy + reassign |
| "invocable action does not exist" | Apex class not in custom PS (auto-generated PS incomplete) | Create custom `{AgentName}_Access` with all `<classAccesses>` (Step 3) |
| "Invalid default_agent_user" | Username typo or user not active | Query Einstein Agent Users, verify exact username + `IsActive = true` |
| Agent runs but returns wrong data | Employee agent using wrong user context | Verify `agent_type` — Service agents use dedicated user, Employee agents use logged-in user |
| `sf org create user` fails | Used in production/sandbox org | Use `sf data create record` instead (Step 1, Option B) |
