Sentinel

A config-driven Salesforce app that surfaces stale, at-risk records across any standard object, with a colorful dashboard, a five-sub-agent Agentforce assistant, one-click triage, and a daily proactive monitor. Zero per-object code.

Ask Sentinel "what should I work on first?", open the Sentinel app for a live health radar, or let the daily monitor push you a digest. Adding a new object to watch is a metadata record, not a code change.

---

⚠️ Before you deploy to a new org

Three things to configure, the deploy will succeed without them, but the agent won't behave correctly until they're set:

1. Agent user. Sentinel ships as an Employee Agent (internal users, no channel setup, no agent user required). If you switch it to a customer-facing Service Agent, edit Sentinel.agent: set agent_type: "AgentforceServiceAgent" and add an access: block with a real Einstein Agent User, publishing with a missing/placeholder user fails with "Internal Error, try again later."
2. Permission set. Assign Sentinel_User to the agent's running user and to end users, otherwise actions return no data or are blocked.
3. Config records. Review the Sentinel_Config.* metadata records so the watched objects, date fields, thresholds, and filters match your org. (The shipped Date_Field__c values must be queryable on that object, e.g. LastActivityDate exists on Opportunity/Lead but not everywhere.)

The same checklist lives at the top of Sentinel.agent and manifest/package.xml.

---

Why it's effective

Most record-hygiene solutions hard-code one query per object. Sentinel inverts that: all object-specific knowledge lives in a Custom Metadata Type, and a single Apex service reads that config to build every query at runtime.

Property	How Sentinel achieves it
No per-object code	One SentinelService serves Case, Opportunity, Lead, Task, and anything else you configure.
No storage overhead	Config ships as Custom Metadata, deploys with the app, packageable, no records to seed.
Change without deploying Apex	Thresholds, date fields, sort, filters, and alert levels are metadata edits.
Reliable agent actions	The agent calls Apex through thin flows (flow://), the pattern Agentforce treats as most reliable.
Governor-safe & injection-safe	Queries are capped and ordered; every identifier from config is validated against an allowlist.

---

Architecture

Conversation ─ Agentforce "Sentinel" Employee Agent
                 └─ router → 5 sub-agents (see below)
Automation ── Daily SentinelMonitor → Chatter digest + threshold notifications
UI ────────── "Sentinel" Lightning app → 7 LWC tabs on one shared design system
Actions ───── 5 flow-fronted invocables: HealthCheck · Rollup · Reassign · FollowUp · Catalog
Logic ─────── SentinelService (shared engine) + focused controllers
Data ──────── Sentinel_Config__mdt (one record per watched object)

Data, Sentinel_Config__mdt

Field	Purpose
Object_API_Name__c	Object to monitor (Case, Opportunity, Lead, Task).
Date_Field__c	Field measuring staleness (LastActivityDate, LastModifiedDate, ActivityDate).
Stale_Days__c	Days before a record counts as at-risk.
Owner_Field__c	Usually OwnerId; resolves Owner.Name automatically.
Priority_Field__c	Optional field to surface (Priority, StageName, Status).
Name_Field__c	Display-name field, Name, or CaseNumber / Subject.
Filter_Criteria__c	Optional extra WHERE fragment (IsClosed = false).
Sort_Field__c / Sort_Direction__c	Optional ordering (defaults to date, oldest first).
Threshold_Alert__c	Daily monitor raises a notification at/above this count.
Is_Active__c	Unchecked configs are skipped by rollups and the monitor.
Label__c	Friendly name shown to users (Active Pipeline).

Four records ship: Active Pipeline (Opportunity), Open Cases (Case), Working Leads (Lead), Open Tasks (Task).

Logic, SentinelService

One engine builds each query from config and returns an enriched report per record: freshness band (Aging / Stale / Critical at 1×/2×/3× the threshold), deep link, resolved owner name, age in days, plus summaries grouped by owner, priority, and band. Every entry point reuses it.

The agent, 5 sub-agents

A hub-and-spoke Employee Agent; the router picks a specialist per utterance:

Sub-agent	Handles	Backing
Record Health	"stale opps?", "check my leads"	Check_Health, Check_All
Focus / Prioritize	"what should I work on first?", "morning briefing"	Check_All scoped to you → ranks → hands off to Triage
Triage	"reassign these", "create follow-ups"	Reassign_Records, Create_Follow_Up, confirms first
Catalog	"what can you check?", "how does staleness work?"	List_Watched_Objects
Coach	"how do I keep my pipeline fresh?"	best-practice advice, then offers real data

Focus and Coach transition into other sub-agents, so one thread can flow check → prioritize → act.

UI, the Sentinel Lightning app

Seven tabs share one "control room" design system: a deep violet header rail, a light data surface, and chart-forward panels.

Tab	Component	Reads
Sentinel	sentinelDashboard	Critical-first hero scoreboard, critical-share gauge, All/Mine scope, a KPI card per watched object (freshness bands, real 7-day trend, expandable oldest records)
Trends	sentinelTrends	Inline-SVG trend lines per object, 7/14/30-day range, improving/worsening badge, expandable data table
Team Rollup	sentinelLeaderboard	Ranked owner scoreboard with position treatment, share of total, search, sort, pagination
Snoozed	sentinelSnoozed	Deferral register with wake-up countdown and row actions
Pipeline Value	sentinelPipelineValue	Amount at risk, critical-share gauge, band split, breakdown by stage
Single-Threaded	sentinelSingleThreaded	Relationship-risk register with severity pills and per-row "why" detail
Activity Gaps	sentinelActivityGaps	Never contacted vs. went cold per object, with a never-worked share gauge

Nine shared LWCs carry the system so nothing is duplicated: sentinelSectionHeader, sentinelStatCard, sentinelStatusBadge, sentinelBandBar, sentinelGauge, sentinelFilterBar, sentinelSkeleton, sentinelEmptyState, sentinelErrorState, plus the sentinelUtils service module. Every view has a skeleton loading state sized to its loaded content, a friendly error state with retry and expandable diagnostics, and an empty state that reads "all clear" as a win. Charts are hand-rolled inline SVG with role="img" labels and a text equivalent, no charting library.

Automation, SentinelMonitor

A scheduled sweep that runs every active config, posts a Chatter digest, and raises a Sentinel Alert custom notification for any config at/above its Threshold_Alert__c.

---

Install

sf org login web --alias my-org --set-default

# Deploy everything (uses the manifest):
sf project deploy start --manifest manifest/package.xml --json
#   …or just: sf project deploy start --source-dir force-app --json

# Grant access, end users AND the agent's running user
sf org assign permset --name Sentinel_User

# Publish & activate the agent (see "Before you deploy" first)
sf agent validate authoring-bundle --api-name Sentinel --json
sf agent publish  authoring-bundle --api-name Sentinel --json
sf agent activate --api-name Sentinel --json

# Schedule the daily monitor (anonymous Apex)
echo "System.schedule('Sentinel Daily Monitor','0 0 7 * * ?', new SentinelMonitor());" \
  | sf apex run

Note: sf project deploy start deploys the agent bundle as an editable draft; sf agent publish compiles it into a running Bot. If the CLI's Einstein authoring endpoint is unreachable, you can Publish → Activate from Agentforce Studio in the browser instead.

---

Using Sentinel in the Salesforce UI

• Dashboard (works immediately): App Launcher → Sentinel, or the Sentinel tab. Drop the sentinelDashboard component on any Home/App/Record page via Lightning App Builder.
• Agent (after publish + activate): in Agentforce Studio → Sentinel → Connections, add the Salesforce (Employee) channel and enable it for Sentinel_User holders. It then appears as the Agentforce assistant panel (sparkle icon, top-right utility bar) on any Lightning page, and optionally in Slack.

---

Demo data

scripts/seedData.apex creates ~60 at-risk records across Opportunity, Lead, and Task by backdating completed activities (which drive LastActivityDate) and setting past ActivityDate on open tasks. Re-run any time:

sf apex run --file scripts/seedData.apex

---

Known limitations

• Ageing Cases depends on the org: Case needs a queryable staleness date. Where LastActivityDate is absent and audit-field writes are disabled, seeded Cases can't be backdated, the Case board reads 0 until cases age naturally. Opportunity/Lead/Task populate fully.
• ownerScope=mine filters on the owner field equalling the running user; use it only where that field references a user.
• Dynamic SOQL is assembled from metadata. Config is developer/admin-controlled and every identifier is allowlist-validated, but treat Filter_Criteria__c as trusted developer input, never end-user input.
• CLI publish endpoint. For some org types the CLI routes sf agent publish to test.api.salesforce.com; if that host is unreachable, publish from Agentforce Studio in the browser instead.

---

Project layout

manifest/package.xml                   # deploy manifest + config checklist
scripts/seedData.apex                  # demo data
force-app/main/default/
├── objects/Sentinel_Config__mdt/      # Custom Metadata Type + 12 fields
├── customMetadata/                    # one record per watched object
├── classes/                           # SentinelService + 5 actions + controller + monitor + tests
├── flows/                             # 5 flow-fronted action targets
├── notificationtypes/                 # Sentinel_Alert custom notification
├── lwc/                               # 7 page components + 9 shared components + sentinelUtils
├── flexipages/  tabs/  applications/  # Sentinel app page, tab, Lightning app
├── permissionsets/                    # Sentinel_User
└── aiAuthoringBundles/Sentinel/       # the Sentinel agent (5 sub-agents)

Screenshots

Captured from a live org at a 1200px viewport.

Dashboard. The control room. The critical count is the one number readable from across the room.	Trends. Real axes, gridlines and labels, with a text alternative for the chart.
Team Rollup. Ranked owners with share of total. No invented deltas.	Activity Gaps. Never contacted and went cold, with the records behind each count.

License

MIT © 2026 VurtuoLabs