CommerceIQFDE OS
Sign in
CommerceIQ · Internalv0.10.0 · retail AI control plane

FDE OS

Operating System & Agent Mesh for Retail AI

The control plane for CommerceIQ forward deployed engineers. Delivery runs as a state machine. The agent mesh is grounded in SOPs and live systems — and every write to a customer system of record still goes through a human.

Web OS
Portfolio, pilots, CRM, builder, evals
Fdee
Slack enterprise engineer on the same brain
REST
Versioned internal API + playground
HITL
Wizards for every mutating step
Sign inOpen appInspect the mesh →

15

Pilot stages

Content Agent clock

HITL

SoR writes

Human executes

T1–T4

Connector tiers

Hard capability ceilings

4

Knowledge scopes

System · SOP · project · personal

3

Mesh channels

Web · Slack · API

RLS

Authz kernel

@commerceiq.ai only

Operating system

Four layers. One FDE bench.

01Kernel

Identity, tenancy, flags, audit, and the capability ceiling. Nothing in the mesh can outrun this layer.

  • Google OAuth · @commerceiq.ai domain gate
  • Role + module RLS (FDE, Head of FDE, CSM, PM, AE)
  • Immutable audit trail on reads, agent turns, HITL
  • Feature flags, API keys, impersonation

02Delivery OS

Engagements run as state machines — not slide decks. Phase, gate, owner, and blocker are first-class.

  • Content Agent 15-stage pilot clock + SLA buckets
  • SOP phases: Pre-flight → DSA → Platform → PIM → CS → verify
  • Programs, staffing, capacity, two-in-a-box pairing
  • UAT plans, business reviews, outcomes / reliance taper

03Agent Mesh

Personas, skills, and tools share one grounded runtime. Query-only to systems of record; drafts for humans.

  • Delivery copilot (web) · Fdee (Slack) · project agents
  • Skills: Fdee brain, pilot delivery, enablement, pulse
  • RAG over SOP, knowledge graph, Gmail / Drive / Calendar
  • Readiness checks against DSA, crawl, PIM — read-only

04Connectors

Tier, not config, decides what is possible. Tier 1/2 have no write path in the codebase.

  • T1 client SoR: PIM, retailer portals — read-only
  • T2 internal SoR: Databricks, DSA, GitHub, S3 — read-only
  • T3 intelligence: CommerceIQ Brain, Content Agent — query
  • T4 collab: Slack, Linear, Google — confirmed write

Agent mesh

Personas and product agents on one runtime

Fdee in Slack, the delivery copilot in the OS, and engagement-bound project agents share tools, skills, and the knowledge graph. Retail AI systems (Content Agent, Brain) join as Tier 3 query nodes — never as unsupervised writers.

FDE OS kernel

online

Control plane

Auth · RLS · audit · flags

Delivery copilot

active

Web mesh node

SOP-grounded next-step + HITL drafts

Fdee

active

Slack mesh node

Portfolio pulse · success · growth

Project agents

assignable

Engagement-bound

Scoped tools + knowledge per pilot

Content Agent

ready

Tier 3 query

Catalog / readiness intelligence

CIQ Brain

ready

Tier 3 query

Account context & recommendations

Grounded, not chatty

Procedural answers must hit SOP / playbook retrieval. On a miss the mesh defers — it does not invent steps, tables, or thresholds.

Look, don’t touch

Readiness SQL, crawl coverage, PIM match, and Brain queries are read-only. Mutating work is emitted as a sectioned HITL wizard.

One brain, many surfaces

Web, Slack, and API share the same runtime, skills catalog, and audit log. Fdee is a persona on the mesh — not a second stack.

Skills are attachable

Pilot delivery, enablement ops, executive pulse, success storytelling, and growth content bind tools to functions and channels.

Module catalog

The OS is the system of record for FDE work

Delivery

  • HomeExec pulse, assignments, at-risk pilots
  • PortfolioEvery engagement × phase × owner × blocker
  • EnablementRASCI, 30-day path, capacity, briefs
  • ProgramsLive two-in-a-box KPI programs + SWAT
  • OutcomesTTV, reliance taper, 70/30 work split
  • StaffingFDE load, allocation, overallocation flags

CRM

  • AccountsTier, products, health, paired CSM/AE
  • ContactsStakeholders, champions, economic buyers
  • CampaignsEnrollment and performance tracking
  • TasksPhase-linked work with Linear / JIRA refs
  • SupportPublic intake → FDE inbox
  • SearchKnowledge graph + agent memory

Builder

  • LibraryReusable solution assets
  • PlaybooksCanonical SOP as executable content
  • CookbooksRepeatable implementation recipes
  • ChecklistsGated step lists with owners
  • EvalsGolden sets + meta-eval release gate
  • WizardsHITL command / payload / PR drafts

Business

  • UATVerification → recorded customer sign-off
  • ReviewsBusiness-review cadence per account
  • OpportunitiesPipeline tied to FDE fit
  • SuccessCase studies, testimonials, proof
  • GrowthHackathon, conferences, webinars, blogs
  • QualifySix-gate FDE fit test

Mesh & tools

  • AgentSpacePlatform agents + per-project studio
  • FDE AgentDelivery copilot — SOP, next step, HITL
  • FdeeSlack enterprise engineer on the same brain
  • PlaygroundAgent / API console
  • AI SkillsPersona, tools, channels, status
  • Control panelRuntime, flags, connector health

Admin kernel

  • UsersRoles, managers, module access
  • FlagsNav and capability switches
  • IntegrationsTiered OAuth / token vault
  • API keysMachine access to the OS
  • Audit logWho did what, to which customer
  • FeedbackBugs, enhancements, review queue

Retail AI products

Multi-product bench, Content Agent first

Content Agent

Full SOP

Only product with the full onboarding state machine in v1 — 15-stage pilot, evals, UAT, HITL config drafts.

Amazon Copilot

Delivery

Portfolio + generic phase/checklist until product SOP is encoded. Same kernel, same mesh.

Digital Shelf

Delivery

DSO engagements as first-class projects with CRM, staffing, and outcomes attached.

Retail Media

Delivery

RMM programs tracked beside Copilot and Content Agent on the same FDE bench.

Safety stance

The mesh can look. It cannot touch.

  • Structural, not policy

    Tier 1/2 connectors do not implement write methods. CI fails the build if they grow one.

  • Human disposes

    The mesh proposes commands, payloads, and PR diffs. An FDE runs them and records the outcome.

  • Grounding required

    No invented SOP steps. Retrieval miss → escalate, don’t hallucinate a runbook.

  • Everything is logged

    Agent turns, tool calls, readiness reads, HITL execution, and evals write an immutable trail.

Open the OS

@commerceiq.ai accounts only. Sessions are refused for any other domain before they are issued.

Sign inOpen app