The Agentic Operating System for E-Commerce

Architecture: 6-Layer OS

Each layer depends only on the one below. Ship L1 + L2 + one agent without building the full stack.

Data Architecture

Four layers from raw ingestion to decision memory. Each layer adds proprietary value.

Canonical Schema (16 Entities)

ChannelListing is the Rosetta Stone — every daily table FKs to it. MasterSKU is at variant level, not parent.

6 Throttle Signals

Every active SKU is scored on 6 signals. Computed deterministically in L2.5 (Evidence Layer) on cron. Consumed by L2 junior agents for execution decisions and by L1 senior managers for strategic oversight.

Knowledge Architecture

Every diagnosis has two parts: the process (universal) and the interpretation (domain-specific). Terminal6 represents both as structured, composable layers that are assembled at runtime.

How Knowledge Composes at Runtime

Category cards and region cards compose at L1 (senior manager level) — where domain diagnosis happens. L0 doesn't need phone-lifecycle interpretation. L2 doesn't need salary-cycle context. L1 is where expertise matters.

This diagnosis is impossible without composing all layers. But the composition happens at the right level (L1), not everywhere. L0 triages without it. L2 executes without it.

Agent Design: Hierarchical Team

The agent architecture mirrors a real e-commerce team. Junior agents are specialists who monitor and execute. Senior managers diagnose within their domain and resolve conflicts between their reports. The Brand Head (human operator) makes strategic decisions.

The Hierarchy

Five Senior Managers, Each With Specialists

Why Hierarchical?

Hierarchical Conflict Resolution

The Autonomy Flywheel

The system earns autonomy by encoding operator decisions as directives. Each strategic call the operator makes gets stored, so the same situation auto-resolves next time.

Day 1: No directives. Everything escalates. Day 30: 2–3 novel situations/day. Day 90: System proposes new directives from patterns. Operator evolves from manager → strategist.

Brand Directives: Living Strategy

Operator intent from meetings and ad-hoc decisions, captured as structured directives with priority stacking.

Key principle: Directives give intent. The Policy Engine's hard constraints are never overridden. A "go aggressive" directive cannot push margin below the floor.

Context Assembly

Each level of the hierarchy sees different data at different resolution. The harness assembles tailored context per agent call.

All levels include brand context (~2K tokens) + channel knowledge (~3K) + evidence brief (~1K). Total budget: ~10–20K per call. The agent never sees all 50K SKUs — it sees the right 5–10 with deep context.

Policy Engine

Every agent action passes through the policy gate. Hard guardrails that agents cannot bypass.

Three-Level Skills

Skills match the hierarchy. Each level receives findings from above and produces output for the level below AND a summary going up.

Two L0 roles

Routing: Declared in Skill, Activated by Evidence, Dispatched by Harness

Every investigate.md declares an Input Metrics table mapping owned metrics → downstream agents. When a metric deviates, the agent checks which input metrics also deviated and invokes the owning agent. The harness assembles context. Fully auditable.

Decision Flow

The harness walks the tree: top-down for investigation, bottom-up for summary.

Typical call volume per trigger: 1 Haiku (L0) + 1–2 Sonnet (L1) + 1–3 Sonnet (L2) = 3–6 LLM calls. Not every trigger activates all branches.

Monitoring: Reactive + Proactive

Two directions of information flow. Reactive (top-down): outcome degraded, investigate why. Proactive (bottom-up): input changed, act before the outcome degrades. Both use the same harness, routing, and policy engine.

Reactive: Two-Tier Outcome Monitoring

Static thresholds don't work — validated empirically. Replaced with self-calibrating detection + LLM triage.

Full investigation skill (Sonnet) only runs when Tier 2 says "yes" — expensive but rare (2–3 situations/week, not 30/day).

Proactive: Input Monitoring (Bottom-Up)

Don't wait for revenue to drop. L2 agents monitor the inputs to revenue — campaigns, listings, inventory, pricing, delivery — and act before the impact hits.

Proactive Alert Flow

The Morning Briefing Combines Both

This is the difference between a recommendation engine and an operating system. Without proactive alerts: "Revenue dropped 30% over the weekend — why?!" With proactive alerts: "Campaign X was suspended and auto-restarted at 8:12pm. Revenue impact: <₹2K. No action needed."

Conversational Interface

Alerts and briefings are push-based (system → operator). But operators also need to pull — ask questions, explore data, make sense of what they're seeing. The chat is the missing piece between the morning briefing and the next day's briefing.

Three modes of interaction

One interface, four modes

The operator talks to one interface — Terminal6. Behind it, the system detects intent and activates the right mode. The operator never chooses a mode — the system figures it out.

Strategy mode: planning with skills

When the operator asks for a plan, the system fans out to multiple L1 domain experts, then synthesizes their inputs into a structured deliverable. Same hierarchy, same agents — used in planning mode instead of monitoring mode.

Planning skills (chief_of_staff + brand_head)

Each skill is a structured planning document — same structure as investigation skills (input → reasoning → routes → output) but triggered by operator request, not anomalies, and producing structured deliverables.

Conversation flow: from question to directive

Same hierarchy, same routing, same policy engine. The only difference: the trigger is a chat message instead of an anomaly or a cron job.

Data grounding (the hard part)

Every answer must trace to a real database query. The LLM decides what to query; tools execute it; the LLM formats the answer. Data is always real, never generated.

Phased rollout

Why this matters for the business

Engagement is the wedge. The operator starts by asking questions (zero setup, zero trust required). Gets useful, grounded answers. Builds trust. Enables alerts. Enables autonomous actions. Chat is how you earn the right to act. Daily active usage is the leading indicator of expansion from intelligence → execution → full OS. Every conversation is also a directive source — accelerating the autonomy flywheel.

LLM Strategy

Context assembly, not model training. Skills, brand context, category cards, and decision history are injected at runtime. No fine-tuning needed.