Fleet

Continuous agent QA and behavioral drift detection for Agentforce.

Fleet answers one question on a loop: is this agent still behaving the way we approved? It captures every agent turn, replays a curated golden set against each monitored agent, scores the deviation from a blessed baseline, attributes any drift to the change that caused it (a deploy, a KB edit, a model rollforward, a prompt edit), and drives remediation with an approval trail. It is Salesforce-native - a React UI Bundle running on Multi-Framework, with all logic in Apex. No external services.

The loop: Capture → Calibrate → Judge → Attribute → Remediate.

Architecture at a glance

• Objects - Fleet_Agent__c, Golden_Case__c, Assertion__c, Calibration_Run__c, Case_Result__c, Deviation_Finding__c, Remediation__c, Change_Event__c, Fleet_User_Preference__c.
• Capture layer - the Fleet_Turn__e platform event taps every agent turn; a trigger archives it into the Fleet_Trace__b big object.
• Admin config - thresholds, detectors, severity policy, assertion types, and console views live in Custom Metadata (Fleet_*__mdt), changeable without a deploy.
• UI - the fleetUi UI Bundle (React 18 + Vite + Tailwind), surfaced as the Fleet Lightning app (uiBundle ref c__fleetUi). Internal route/feature namespace is agents to avoid colliding with the product name.
• Identity - teal #0F766E primary, sky #0EA5E9 accent, rose severity; Space Grotesk headings, Inter body, JetBrains Mono metrics.

Prerequisites

• Node.js v22+
• Salesforce CLI v2.130.7+ (ships the UI Bundle plugin) - check sf --version, update with sf update.
• Target org on Summer '26 or later (Multi-Framework is GA and on by default).

Deploy (per CONTRACT §13)

bash scripts/auth.sh                 # authorize a target org, alias "fleet"
bash scripts/deploy.sh               # build the UI, deploy force-app, assign Fleet_Administrator
bash scripts/deploy.sh --check       # validate-only (dry run)
bash scripts/deploy.sh --no-ui       # deploy metadata only, skip the Vite build
sf agent validate authoring-bundle --api-name Fleet_Curator
sf org open --path lightning/app/Fleet

Deployment order matters: objects → Apex → Flows → agent bundles. The agents reference Apex and Flows by name, so they publish last.

The root package.json also exposes helpers that delegate into the bundle:

npm run ui:install     # install UI Bundle dependencies
npm run ui:dev         # run the console standalone with the mock data adapter
npm run ui:build       # produce the dist/ the UI Bundle host serves
npm run apex:test      # sf apex run test with coverage

Standalone UI development

The UI Bundle runs without an org. getRepositories() switches on VITE_DATA_MODE (default mock), so the whole console can be built and demoed against seeded mock data before any Apex exists. npm run ui:dev serves it from / (no SFDC_ENV, so the router basename is undefined); inside the org it is served from /app/c__fleetUi.

Packaging note

Multi-Framework apps cannot ship in a managed package yet. Ship the agents, objects, Apex, Flows, and CMDT as the managed package, and the UI Bundle as an unlocked package installed alongside. Collapse them when managed-package support lands.

Testing

The Apex suite and the UI suite run independently.

npm run apex:test                              # Apex, with coverage
cd force-app/main/default/uiBundles/fleetUi
npm test                                       # Vitest, jsdom
npm run lint                                   # ESLint, zero-warning policy
npm run build                                  # tsc -b && vite build

Coverage targets are set in CONTRACT.md §13. The scoring and drift services carry the highest bar because a wrong score silently mis-grades an agent, which is worse than a crash: a crash gets noticed.

Testing note on big objects

Fleet_Trace__b cannot be written in an Apex test context at all, and Database.insertImmediate behaves like a callout, so it fails after ordinary DML in the same transaction. FleetTraceService therefore writes through a TraceWriter seam with an in-memory implementation under test. Do not collapse that indirection; the tests cannot run without it.

Repository layout

fleet/
  CONTRACT.md                     single source of truth for shared names and shapes
  force-app/main/default/
    classes/                      Apex services, invocable actions, tests
    objects/                      custom objects, big object, platform event, CMDT
    permissionsets/               Fleet_Administrator, Operator, Approver, Viewer, Integration
    genAiPlannerBundles/          Agentforce agent definitions
    uiBundles/fleetUi/            React console (source, tests, Vite config)
  scripts/                        auth, deploy, seed
  docs/                           design notes

Screenshots

Captured from a live org at a 1200px viewport.

Dashboard. Is every monitored agent still behaving the way we approved? Trueness against threshold leads, because that is the number that decides whether anyone needs to act.	Agents. Every monitored agent with its trueness, blessed version, quarantine state and last calibration.
Drift. Trueness over 72 hours for one agent against the rest of the fleet, with change events marked on the same timeline so a drop lines up with what caused it.	Findings. Detected drift, the change it was attributed to with a confidence score, and the action taken. Rollback and hold are one click from the finding.

License

MIT. Copyright (c) 2026 VurtuoLabs. See LICENSE.