Fleet answers one question on a loop: is this agent still behaving the way we approved? It captures every agent turn through a platform event, 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 knowledge-base edit, a model rollforward, a prompt edit), and drives remediation with an approval trail. The loop is Capture, Calibrate, Judge, Attribute, Remediate. It is Salesforce-native: a React UI Bundle running on Multi-Framework with all logic in Apex, and no external services.
Key features
Capture layer: the Fleet_Turn__e platform event taps every agent turn and a trigger archives it into the Fleet_Trace__b big object
Golden-set replay: curated cases with assertions, replayed against each monitored agent and scored for trueness against a blessed baseline
Drift attribution: deviations tied to the change that caused them (deploy, KB edit, model rollforward, prompt edit) with a confidence score
Remediation with an approval trail: rollback and hold are one click from a finding, with quarantine state tracked per agent
Admin config in Custom Metadata (Fleet_*__mdt): thresholds, detectors, severity policy, assertion types and console views, changeable without a deploy
React 18, Vite and Tailwind console surfaced as the Fleet Lightning app, with 8 feature routes and a mock data adapter for standalone development
Five least-privilege permission sets: Administrator, Operator, Approver, Viewer and Integration
Use cases
QA lead checks each morning whether every monitored agent is still inside its trueness threshold
Release manager lines a trueness drop up against the deploy or prompt edit marked on the same timeline
Operator quarantines a drifting agent and rolls it back to the blessed version straight from the finding
Approver signs off on a remediation before it reaches production, leaving an approval trail
Admin tunes drift thresholds and severity policy in Custom Metadata without a deploy
Team builds and demos the console standalone against seeded mock data before any Apex exists
Industries
AI Operations, Quality Assurance, Platform Engineering
Business need
Continuous assurance that a deployed Agentforce agent still behaves the way it was approved to behave, with drift caught and attributed to its cause before customers notice
Requires
Fleet_Administrator permission set; Agentforce enabled for agent publish and activate; target org on Summer '26 or later (Multi-Framework GA); Node.js v22+ and Salesforce CLI v2.130.7+ to build and deploy the UI Bundle
Compatible with
Salesforce Agentforce agents (genAiPlannerBundles); Salesforce Multi-Framework UI Bundles; Custom Metadata driven configuration. No external or third-party services.
Support level
Standard
Screenshots
Product details
Compatibility
API version
Salesforce API v67.0. Requires Summer '26 or later because the UI Bundle runs on Multi-Framework, which is GA and on by default from that release. Lightning only.
Integrations required
Salesforce CLI v2.130.7+ (ships the UI Bundle plugin); Node.js v22+ for the Vite build; an Agentforce-enabled org for agent publish and activate. No external services and no callouts.
Setup time
30 to 60 minutes. scripts/auth.sh and scripts/deploy.sh handle authorize, UI build, deploy and permission set assignment; the variable time goes into publishing the agent bundles and curating the first golden set.
App details
Languages
English
Package type
Metadata (Unmanaged). Multi-Framework apps cannot ship in a managed package yet, so the intended split is a managed package for agents, objects, Apex, Flows and CMDT, with the UI Bundle as an unlocked package installed alongside.
Edition support
Requires an Agentforce-enabled org on Summer '26 or later. Uses one big object and one platform event; no Data Cloud required.
Mobile compatible
Desktop-oriented. The console is a React UI Bundle Lightning app, built and screenshotted at a 1200px viewport; mobile support is not claimed in the repo.
Security
No data leaves the org: 100% native Apex, Flows and a React UI Bundle, with no external callouts. Five least-privilege permission sets separate Administrator, Operator, Approver, Viewer and Integration. Big object writes route through a TraceWriter seam rather than being written directly, and deployment order is enforced objects, Apex, Flows, then agent bundles.