The four agents
| Agent | What it produces | Audience |
|---|---|---|
| L1 — Category Brief | What the market values: pain points, growth levers, audience segments | Brand strategy, category planning |
| L2 — Perception Report | How the brand's products perform: feature sentiment, leaderboards, trajectories | CMO, product marketing |
| L3 — Situation Analysis | What to do: priorities, gaps, recommended actions per product | Ecommerce, brand managers |
| L4 — Content Generation | Ship: Amazon listings, brand PDPs, retailer PDPs, social copy | Content team, agencies |
Each agent reads from the pillar tables built upstream. None of them touch raw text. This is what makes them fast, cheap, and reproducible.
Why a chain, not a single agent
A single "do everything" LLM call on raw snippets would be:
- Slow — 100K+ snippets per category is too much context
- Expensive — every output regenerates the same intermediate analysis
- Inconsistent — same question gives different answers across runs
The chain solves all three:
- L1 produces a stable category model. Cached.
- L2 produces a stable perception model on top of L1. Cached.
- L3 reads both. Cheap.
- L4 reads L3. Cheap and ship-ready.
Each layer has a single responsibility and a stable output schema. Downstream consumers don't break when upstream agents are tweaked.
What each agent reads and writes
L1 Category Brief reads:
rag_snippets(perception)distinctive_keywords_segment(demand)keyword_search_volume(demand)
Writes: structured category brief — pain points, growth levers, segment maps, top properties.
L2 Perception Report reads:
sentiment_trajectories(per product × property × period)property_leaderboard(cross-product ranking)- L1 output
Writes: per-product perception summary with feature/benefit/use_case sentiment, trajectories, competitive position.
L3 Situation Analysis reads:
- L2 output
brand_gaps(what brand says vs what market says)product_edges(closest competitors per product)
Writes: priority-ordered action plan per product, with evidence, gap classification, and recommended next move.
L4 Content Generation reads:
- L3 output
asin_content(current Amazon listing)- L2 perception (claims to emphasise)
Writes: full Amazon listing (title, bullets, A+ content), brand PDP copy, retailer PDP copy, social posts — ship-ready.
Model selection by agent
| Agent | Model | Why |
|---|---|---|
| L1 | Claude Sonnet | Needs reasoning about market dynamics |
| L2 | Claude Sonnet | Multi-product synthesis |
| L3 | Claude Sonnet | Competitive logic |
| L4 | Claude Sonnet for copy, GPT-4o-mini for title verification | Sonnet for creativity, mini for rule-checking |
Per-run cost: ~$0.50 for L1-L3, ~$2 for L4 across a full ASIN portfolio. The expensive step is the upstream enrichment — by the time strategy agents run, the heavy lift is done.
Why this matters
The strategy chain is the delivery surface of Theia. It's what a CMO actually reads, what a brand manager actually ships from, what a research firm puts in front of a client.
Behind it is the intelligence repository. In front of it is the deck, the listing, the page, the brief. The chain is the bridge.