A hundred AI agents argued over a market.We gave them a Myanmar conglomerate, a coffee category, and eight countries. They came back with a brand, a budget, and a year of launch plans. No humans in the loop.
DebateAI runs an autonomous multi-agent debate: advocate, red team, CFO, judge. Every claim gets attacked, every market scored, every campaign put through a tournament. What you are about to scroll is the actual deck it produced.
Benchmark Food EmpireS$76.2M from CIS8 markets debatedWinner Philippines 7.455 campaigns ranked$13M playbookROAS 1.8×25,000 stores by M12brand LaKás Kapemarket $1.96B
Prologue
No strategist wrote this deck. A hundred agentsargued it.
CDSG, a Myanmar holding, had never sold a sachet of coffee. We handed the brief to an adversarial multi-agent system and stepped back. The market it picked, the brand it named, the film it cut: every decision below is the work of the room, scored and ruled in twelve minutes.
The Debate · June 2026 · for CDSG Holdings
01The Brief
A Myanmar holding company. A coffee market it had never sold in.
CDSG wanted into the Southeast Asian instant coffee market. The mandate: pick the right country, build a local consumer brand, and design the full launch, all while keeping the Myanmar parent invisible to shoppers. We handed the brief to the agents and stepped back.
$0M
PH Market Size
instant coffee, 2024-Q1
0%
5-Year CAGR
high-growth category
0kg
Coffee / Capita
annual consumption
0
Sari-Sari Stores
last-mile network
ClientCDSG Holdings Myanmar conglomerate, origin hidden from consumers
CandidatesPhilippines, Vietnam, Thailand, Malaysia, Cambodia, Indonesia, Bangladesh, Laos
Cost ModelBYOK. Runs on the user's own Google AI key. $0.10 to $0.40 per full analysis.
OutputMarket pick, brand, creative, $13M budget, 12-month calendar, KPIs, risk plan
Runtime~12 minutes fully autonomous, no human in the loop
02The System
Four phases. A hundred agents. Each one argues, then hands off.
A Python MasterOrchestrator drives the phases in sequence. Inside each phase, dozens of specialist agents run in parallel: building cases, attacking them, scoring the result. Output is validated against strict JSON schemas before it moves on.
PHASE 01
12 agents
Intelligence
Decode 25 years of Food Empire reports into a transferable success formula.
PHASE 02
48 agents
Market Debate
Advocate, red team, CFO and judge fight over eight markets. Weakest cut.
PHASE 03
32 agents
Tournament
Five campaigns generated, critiqued by specialists, scored and ranked.
PHASE 04
12 agents
Playbook
Refine, validate unit economics, name the brand, ship the board deck.
01
Intelligence Phase
Food Empire, decoded in numbers.
Twelve ingestion agents read Food Empire's annual reports from 1992 to 2002. A historian segmented the eras, deep-dive agents analysed each one, and a synthesis agent (checked by a professor critic) wrote the brief. Eighteen evidence citations, zero human curation.
S$0M
CIS Revenue by 2001
90.5% of turnover
0%
Net Profit Jump
single year, 2001
0%
Revenue Growth
from localisation
0
Sachets / Year
60–70% utilisation
MASTER THESIS
Food Empire pulled S$76.2M, ninety percent of all revenue, out of the CIS region by 2001 through fierce localisation and vertical integration, lifting net profit 133 percent in a single year. Then it diversified into snacks, frozen food and seafood, and reversed all of it in 2002. The lesson: focused, phased expansion beats broad simultaneous growth.
CIS was 90.5% of revenue before any diversification. CDSG should dominate one or two markets before expanding. Confidence: HIGH.
Lesson 02 · Vertical Integration
1.0B sachets a year at 60–70% utilisation drove the 133% profit jump. Own the supply chain. Confidence: HIGH.
Lesson 03 · Localised Branding
Wet sampling plus local event sponsorship (Ukrainian group ALIBI) drove 42.4% growth. Confidence: HIGH.
Caution 01 · Concentration
Ninety percent from one region is a single point of failure. Plan phased diversification from day one.
Caution 02 · Unfocused Sprawl
The 2001 snacks and seafood push was fully reversed in 2002. Hold the core thesis until the position is secure.
Framing for CDSG
Penetrate one or two core markets deeply. Lean on vertical integration and aggressive local branding. Diversify only in phases.
02
Market Debate Phase
Eight markets enter. The CFO kills four.
Forty-eight agents ran at once. For every market: an advocate built the bull case, a red team attacked it, a CFO applied hard kill criteria (entry cost versus budget, payback period, margin), and a convergence judge weighted seven factors into one score. Drag the globe.
Philippines7.45
Vietnam7.05
Thailand7.00
Laos1.00
Malaysia✕
Cambodia✕
Indonesia✕
Bangladesh✕
CFO FINANCIAL SCREEN · ALL 8 MARKETS
Market
Entry Cost
Budget Ratio
Conservative Payback
FX Risk
Verdict
Philippinesclean pass
$5.58M
1.12×
6.0 yrs
HIGH
viable
Vietnam
$4.65M
0.93×
11.79 yrs
HIGH
reinstated
Thailand
$5.64M
1.13×
7.52 yrs
HIGH
reinstated
Laos
$1.20M
0.24×
8.56 yrs
HIGH
reinstated
Indonesia
$5.40M
1.08×
21.0 yrs
HIGH
CFO kill
Bangladesh
$2.34M
0.47×
17.67 yrs
HIGH
CFO kill
Malaysia
$3.70M
0.74×
27.0 yrs
HIGH
CFO kill
Cambodia
entry cost & market size NOT_FOUND, no meaningful assessment
CFO kill
Philippines was the only market under the 7-year payback ceiling (6.0 yrs, conservative), passing clean. Vietnam, Thailand and Laos breached it but were reinstated under a minimum-diversity rule so the judge had a real field. Indonesia, Bangladesh and Malaysia were killed on 21 to 27-year paybacks; Cambodia on missing data.
CONVERGENCE JUDGE · WEIGHTED SCORES
#
Market
Score
Status
Decisive Factor
01
Philippineswinner
7.45
selected
entry feasibility 9/10
02
Vietnam
7.05
contingency
lesson transfer 8/10
03
Thailand
7.00
surviving
mature, niche play
04
Laos
1.00
floor
no evidence presented
SEVEN-FACTOR BREAKDOWN · TOP 3
Philippines7.45growth 8 · feasibility 9 · transfer 9
"With 110M consumers, coffee culture embedded in 800,000 sari-sari stores, and 6.5% GDP growth, the Philippines is a high-growth, instant-coffee-dominant market with viable payback. That last-mile network took Food Empire a decade to build in Russia. Here it already exists."
RedTeamAgent · attack
"The value claim rests on PHP 7.00 being competitive, yet there is no named competitor shelf price, no SKU, no margin analysis. Nescafé and Great Taste hold decades of loyalty. Without a taste benchmark, the unbeatable-value claim is reckless. This is a critical gap."
AdvocateAgent · rebuttal
"Accepted. PHP 7.00 locks behind a competitor price-audit gate before any media. We add a blind taste test against Nescafé, Great Taste and Kopiko before pricing sign-off, and gate the 10,000-outlet claim behind a field audit. These are discovery gates, not launch blockers."
ConvergenceJudgeAgent · verdict
"Philippines: 7.45 out of 10. Strongest unresolved risk: a value-for-money strategy with no identified competitor share or pricing, which makes entry speculative. But entry feasibility (plus 0.40 over Vietnam) and distribution density decide it. Philippines is the primary market. Vietnam is the contingency at 7.05."
DECISION AUDIT · MARKET SELECTION
The system doesn't just pick a winner, it grades its own confidence. Philippines won on score, but the red team logged 5 critical, unresolved attacks against it (the missing competitor price benchmark above all). Net result: winner confidence: LOW and an automatic human review flag before this market gets a real budget. Highest score does not mean rubber-stamped, every winner ships with its own dissent file.
ADJUSTABLE DEBATE DEPTH · MEASURED A/B
A shallow debate rushes to a verdict. So depth is a dial: each agent runs an N-round attack→rebuttal→counter loop inside a single model call — deeper scrutiny, identical API cost (one call, whether N is 3 or 10). We ran the same eight advocate briefs through the red team at depth 3 and depth 10. The deeper pass didn't just talk more, it found more:
Metric (same input, same cost)
Depth 3
Depth 10
Delta
Markets fully pressure-tested
6 (1 truncated)
8 (all clean)
+2
Total attacks generated
42
56
+33%
Critical findings surfaced
30
47
+57%
Avg. finding detail
461 chars
540 chars
+17%
API calls made
1
1
0 extra
At depth 10 the red team escalated from generic gaps to second-order flaws: "HIGH modern-trade access is a commercial illusion," "$5M is a test budget, not penetration spend," and unaddressed Myanmar-origin geopolitical risk. Deeper reasoning, zero extra cost.
03
Campaign Tournament Phase
Five campaigns enter. The safe bet does not win.
Thirty-two agents generated five campaign concepts, each from a different strategic lens. Two specialist critics, one on creative strategy and one on consumer insight, scored every one. The highest creative score (C3, a clean 9 and 9) lost to the most defensible concept (C1).
C1 · WINNER
Ang Lakas ng Araw-Araw
picked
Cultural Truth Platform
8
creative · keep
8
insight · revise
ROAS 1.8×budget $13MBE M20
Turns the invisible daily grind of the working class into collective pride. Won for being the most authentic and the most executionally safe: emotional depth a competitor cannot copy with a price cut.
C3 · top creative score
Ang Kape ng mga Hindi Matatalo
Creative Fame Platform
9
creative · keep
9
insight · keep
ROAS 2.1×budget $15MBE M18
"The Coffee of the Undefeated." Scored highest on pure creativity, but flagged high-risk: it could read as exploitative if execution slipped. It demanded flawless production to avoid backlash.
C2
CDSG: Lakas ng Araw
Commercial Penetration
7
creative · revise
7
insight · revise
ROAS 1.5×budget $15MBE M18
Solid penetration tactics, but the "Lakas" angle sat too close to C1 and C3 to feel distinct. English taglines diluted the cultural authenticity.
C5
Ang Lakas · Seed Playbook
Three-Universe Strategy
5
creative · revise
4
insight · revise
ROAS 2.5×budget $28MBE M15
An ambitious "tipid, sandali lang, lakas" triptych. The audience drifted upmarket, away from the mass price-sensitive core, on a budget nearly double the winner.
C4 · KILLED
Kasosyo sa Kape
kill · 1/10
Trade-Led Growth
1
creative · kill
1
insight · kill
ROAS 2.5×budget $18MBE M18
Strong B2B logic for 1.3M sari-sari owners, but it is a trade strategy, not a consumer campaign. Both critics killed it for missing the brief. Its tactics were salvaged into the final distribution plan.
WHY THE HIGHER SCORE LOST
C3 scored a perfect 9 and 9 on creativity, yet the synthesis chose C1. The logic: C1 carries the same cultural truth with far less execution risk. C3's "undefeated" framing only works with flawless production, and one wrong note reads as exploiting the poor. C1 celebrates strength, not struggle, with zero self-pity, earning loyalty no price advantage can buy, while staying brand-safe at scale.
PROOF GATES BEFORE SCALED SPEND
Gate · Price Audit
Nescafé, Great Taste and Kopiko shelf prices by SKU before media lock.
Gate · Distribution
Hit 50% of key sari-sari clusters in pilot regions within three months.
Gate · Recall
Plus 20% aided recall in pilot regions before national scale.
Gate · Velocity
Sales 15% above new-entrant baseline plus 5% social engagement.
DECISION AUDIT · TOURNAMENT & FINAL VALIDATION
C1 won across 5 candidates and 7 specialist critics, but the dossier still carries 4 open assumptions (price audit, distribution clusters, recall lift, velocity threshold, all unproven until pilot). In Phase 4, the CFO's unit-economics check on the first draft came back FLAWED, triggering an automatic rewrite across 7 refinement roles before the playbook shipped. Every number in the final deck has a paper trail of who challenged it and why.
Intermission · From Strategy to Identity
A campaign won the tournament. It still needed a name to live under.
04The Brand · Separate Swarm
The first run shipped a campaign but never named the consumer brand. So a dedicated five-agent brand swarm convened: a brand strategist, a cultural consultant, a marketing director, a CFO and a creative director. Twelve names debated. One vote. Unanimous.
Consumer Brand · Swarm Consensus 5 / 5
LaKás
LaKás Kape
"Kape ng Lakas, Araw-Araw. Para sa Bayan."
Tagalog for strength. Five characters, instantly trademarkable through the accent mark, and a direct linguistic anchor to the winning campaign "Ang Lakas ng Araw-Araw." The Myanmar parent stays hidden: CDSG keeps ownership, supply chain and distribution behind a local brand built for Filipino working-class identity. Runner-up "Araw Kape" was eliminated.
✓Brand Strategist✓Cultural Consultant✓Marketing Director✓CFO✓Creative Director
#D97706 amber#7C2D12 brown#FCD34D gold
Final Brand & Product Identity
Packshot 3-in-1 Kapeng Timpla sachetsBrand Card "Kape ng Lakas, Araw-Araw."Character Study Mang Eduardo, continuity reference
The identity that shipped in the film · "Matapang at Creamy" · GPT Image & Gemini
04
Final Playbook Phase
The campaign, already shot.
Twelve refinement agents turned the winning concept into a board-ready playbook, then the boards went to camera. These eight frames are lifted straight from the finished film: the pre-dawn open, the dashboard ritual, the first sip, the toast. Drag to spin.
00:04
Pre-Dawn Manila
00:13
Mang Eduardo · Jeepney Driver
00:19
The Ritual · Product Hero
00:27
First Sip
01:00
Maria · Factory Floor
01:18
Aling Nena · Over the Counter
01:28
The Desk Shift
01:44
The Toast
drag to rotate · eight frames from the 2:02 final cut · timecodes on each card
below 1.5× for two months triggers a mandatory creative review
$23.4M
Sales Volume
implied revenue target from ROAS and media spend
40%
Aided Awareness
end of year one, quarterly consumer surveys
10%
Trial Rate
of target audience, survey plus redemption data
70%
distribution coverage
5%
digital engagement
1,000
#LakasNgArawAraw UGC
80%
POS visibility score
RISK RESPONSE PLAN
High
Product cannot live up to the "Lakas" strength promise
Rigorous supply-chain QC, a rapid consumer feedback loop, quarterly blind taste tests against rivals. Emphasise "Tunay na Lakas" in messaging.
High
The killer assumption fails: recognition does not convert to loyalty
Continuous surveys, focus groups and social listening, with a pre-built pivot to balance emotional recognition against explicit functional benefits if loyalty lags.
Med
Price war or counter-messaging from Nestlé and Universal Robina
Monitor pricing and promos. Hold the "Sulit" value line, keep a tactical contingency budget, defend on emotional positioning beyond price.
Med
Distribution gaps in fragmented General Trade
Heavy distributor incentives, dedicated trade-marketing teams, route-optimisation tech, monthly field audits for POS visibility and stock.
A strategy is a document. This is what it looks like at twenty-four frames a second.
05The Film · Final Cut
The storyboard, now rolling.
Every board above went through image-to-video production: twelve AI shots cut to "Bukang Liwayway", a Tagalog voiceover timed scene by scene, sidechain-ducked music and a LaKás end card. This is the full master, rendered in 8K and delivered here at 1080p.
Ang Lakas ng Araw-Araw · LaKás Kape02:02 · Tagalog VO · Sound on
A Python orchestrator coordinates a hundred agents calling Gemini Flash 2.5. Each produces structured JSON, passes a quality gate, and hands off. You stream the whole debate live, and it runs on your key, so we store nothing.
01
MasterOrchestrator
One Python class drives all four phases. A state machine tracks completion; asyncio pools run agents in parallel within each phase. A stdout redirector streams every line to the browser as Server-Sent Events.
02
Schema-Validated Output
Every agent output is validated against a strict JSON schema. Hallucinated or generic output triggers a retry, up to three, before the quality gate escalates. json_repair recovers malformed responses.
03
BYOK · Zero Server Cost
Users bring their own Google AI Studio key. It lives in browser localStorage only, passed per request, never persisted. A full hundred-agent analysis costs the user between ten and forty cents. The platform pays nothing.
04
Live SSE Streaming
FastAPI on Railway streams the debate as it happens (phase transitions, agent names, findings) to a Next.js 14 front end. No WebSocket complexity, no serverless timeout. Upload your own PDFs to ground phase one.
STACK
gemini-2.5-flash reasoning
imagen-4.0 TVC frames
veo-3.0 motion clips
FastAPI backend + SSE
Next.js 14 + Clerk auth
Railway backend
Vercel front end
asyncio parallel pools
pypdf ingestion
ENGINEERING LOG · REAL BUGS, REAL FIXES
›
Veo 3.0 returned video_bytes = None; it was a download URI, not binary
urlretrieve(uri + "&key=")
›
duration_seconds=5 rejected as out of bounds despite being in range
omit param, use default
›
Imagen method mismatch: generate_image vs the real plural
generate_images()
›
Windows emoji in print() threw UnicodeEncodeError on CP1254 during deploy
ASCII [OK] / [FAIL]
›
Railway: ModuleNotFoundError: agents, only the API folder deployed
bundle agents/ config/
›
localStorage is not defined on the Next.js SSR build
read inside useEffect
›
159MB Railway upload timeout from report PDFs in the bundle
.railwayignore data/
Epilogue
Twelve minutes. One market, one brand, one finished film.
The room left behind a $13M playbook and a launch-ready campaign. It is still open. The next brief is whichever one you bring to it.
D·AI
Sealed by the Debate · LaKás Kape · June 2026 Gemini Flash 2.5 · 100 agents · 12 minutes
RUN IT ON YOUR OWN BRIEF
Your company. Your markets. Your strategy in twelve minutes.
Bring a free Google AI Studio key. Describe your company, pick target markets, upload your docs, and watch a hundred agents debate it live. A board-ready report at the end.
