Social Seeding
Google for Startups · AI Agents Challenge Track 3 · Refactor 22-agent ADK fleet on Vertex AI A2A v0.3

Influencer marketing still means a human runs every campaign by hand.

We refactored that into an agent that is the operator — source, vet, outreach, ship, content, performance — and the buyer pays only for performance: a penny per measured view. Not a report you buy. Views you delivered.

See the business model → The fixture proof
the pricing flip
CPM → VPM
V = View. Billed per measured view delivered — not per impression, not per report.
headline price
$0.01 / measured view
A penny per delivered view. Priced agent-to-agent over the AP2 mandate chain — the operator signs the Intent gate.
The loop · one brief, six stages, your gates

One brief wakes the fleet. It runs the loop and halts at the gates you keep.

A 22-agent ADK fleet is defined; three are wired into this demo workflow. The stage bar advances on its own, then stops at a human approval gate before anything leaves the building.

routed 01Source Creator search over the TikTok corpus via the A2A connector. auto
routed 02Vet Rank + score fit; brand-safety screen on each candidate. approve the shortlist
routed 03Outreach Drafts personalized messages; nothing sends until cleared. approve the send
04Ship Brief, assets, and logistics to the confirmed creators. approve logistics
05Content Posts go live; the loop tracks each placement. auto
06Performance Measured views roll up — the metered unit the buyer pays for. auto
human approval gate — always_ask by default; the human-in-the-loop is the product wired into the demo workflow (3 of 22 defined agents)
The business model · we sell performance, not reports

CPM ✗ → VPM ✓. A penny per measured view, billed agent-to-agent.

the pricing flip
CPM → VPM

$0.01 / measured view — a $10 effective CPM, which we cite only as an in-band benchmark (TikTok CPMs run $4.8–$13.26), never as the headline. The headline is simple: pay only for views you delivered.

all-in gross margin · Model B
~29%

On Model B (D28) we bill the brand one all-in price and pay the creator commission out of revenue. After creator pay, FX, dispute, CS and our ~$7.40 GCP cost-to-serve, the board-grade figure is ~29% all-in gross margin. This is the only figure we call "margin."

AP2 mandate chain · carried agent-to-agent (Agent Payments Protocol)
operator signs
step 1
Intent
→
step 2
Cart
→
step 3
Payment
The operator signs the Intent Mandate (D27) — the chain always requires human approval. The Cart + Payment binding and verifier are code-verified; VPM metering (Apigee X → Pub/Sub → BigQuery) and production billing are designed, not yet live. No automatic charge, no "no-human" billing — and we say so.
Traditional agency · management layer $2,400 · ~45 hrs
→
Social Seeding · same 20-creator campaign $7.40 · ~2 review-hrs
That's a 99.7% cut in the management-LABOR cost line ($2,400 management fee → $7.40 compute; ~43 human-hours saved). This is a labor-cost reduction, never a margin. Creator media spend is identical pass-through in both columns.
Proof · the report stage humans abandon, regenerated

16 posts. 59,498 views. ~7.93% engagement.

Fixture, not a live customer

The reporting stage that human operators abandon is exactly the stage the fleet regenerates — the in-app autopilot report from the wooriliu campaign fixture.

16
posts tracked across the campaign
59,498
views — the metered unit the VPM model bills against
~7.93%
engagement rate across the placements
wooriliu is a fixture seeded from a real campaign's shape — not a live customer. The 16-post / 59,498-view / 7.93% figures are the autopilot report rendered from that fixture, shown precisely so the on-screen number matches the code, not the live /share page (which reads 0 verified / ER 0 — a backend quota artifact).
71.4%
The same MVP, hardened. A creator negotiated mid-reply and the agent used to stall; one deterministic rule fixed it. 71.4% on an unseen adversarial holdout (10/14) — and we keep the 4 misses on screen, not tuned away.
The refactor · Google-standard agent stack, our components ringed

A shipped MVP, refactored into a listable enterprise agent.

orange ring = what we run on Google Cloud · everything else is the standard option in the layer
phase 01

Build

The shipped TikTok influencer-campaign MVP — the loop that already runs.

phase 02

Optimize

One deterministic rule on the negotiation miss — kept only as honest reliability evidence (71.4% holdout).

phase 03 · leads

Refactor

Onto a Google-standard agent stack — listable, discoverable, callable by any enterprise agent over A2A.

LayerRuntime
oursCloud Runserverless · scales to zero
oursVertex AI Agent Enginea representative agent deployed
GKEmanaged Kubernetes
Compute EngineVMs
LayerModels
oursModel GardenGemini 3.5-flash · 3.1-flash-lite
Gemini APIdirect
Imagenimage gen
Open / partnerGemma · 3P
LayerOrchestration
oursADKSequentialAgent: searcher → ranker, each a typed function
LangGraphgraph runtime
CrewAIrole crews
Custom loopReAct
LayerInterop
oursA2A v0.3discover + call · message:send
MCPtool context
OpenAPIREST tools
gRPCservice mesh
LayerSecurity
oursCloud KMS Agent Identitysigned, unique per agent
oursModel Armoron the A2A ingress
IAMaccess control
VPC-SCperimeter
LayerMemory
oursMemory Bankagent recall
Vertex Vector Searchembeddings
Firestoresession state
BigQueryanalytics store
LayerDistribution
oursGemini EnterpriseAgent Gallery · registered + ENABLED
Cloud Marketplacelisting prepped, not listed
AgentSpaceworkspace surface
Direct APIembed
Migrated onto Google Cloud: Cloud Run, with a single representative agent on Vertex AI Agent Engine (not the full 22-agent fleet). Models route through Model Garden; identity is signed in Cloud KMS; Model Armor sits on the A2A ingress; Memory Bank backs recall. Reached over A2A v0.3 so any enterprise agent can discover and call it — registered and ENABLED in a Gemini Enterprise Agent Gallery. Assistant→agent routing is the Google-gated next step; the gallery does not yet answer from our agent.
The wedge · a regional block becomes a distribution path

Korea is excluded from Marketplace payment regions. So we ship over A2A.

the constraint (D2)

Google Cloud Marketplace's payment regions exclude Korea.

A Korean-incorporated entity can't take the direct paid-listing rail without a foreign sub-entity — a 6–12-month legal/banking project. The Marketplace listing is prepped, not listed.

the wedge (D3)

Publish over A2A v0.3 instead — discover and call, no Marketplace billing rail.

Register the agent so Gemini Enterprise customers reach the plan_creator_search skill directly, meter per call, invoice under Korean tax law. A forkable BUSL-1.1 + Apache-2.0 pattern any non-payment-region startup can copy.

Other agents don't use ours — they hire it.
A regional exclusion becomes an open-source distribution wedge — and a second revenue line alongside the direct SaaS tenancy. Korea is APAC, so the APAC Regional prize is in range beside the Grand Prize aim (D50).
Honest scope · what is live vs designed vs prepped

The credibility play: we say exactly what runs and what doesn't.

Full boundaries in scripts/demo/submission/HONEST-SCOPE.md. The figures on this page use BUSINESS-CASE.md consistently — $2,400 / ~45 hrs (agency) → $7.40 / ~2 review-hrs (us); ~29% is the only number called "margin"; 99.7% is the management-labor cost reduction, never margin.

designedVPM metering & AP2 billing are designed (Apigee X → Pub/Sub → BigQuery), not yet live in production.
intent onlyAP2: the operator signs the Intent Mandate (D27); the Cart + Payment chain is code-verified; production billing is the roadmap. Never "no human" / "automatic billing."
fixturewooriliu is a fixture seeded from a real campaign's shape — not a live customer. 16 / 59,498 / 7.93% are the autopilot report from that fixture.
enabledGemini Enterprise is registered + ENABLED in the Agent Gallery; assistant→agent routing is the Google-gated next step — it does not yet answer from our agent.
1 of 22Vertex AI Agent Engine runs a single representative agent — not the full 22-agent fleet.
3 of 22The fleet = 22 agents defined; 3 are wired into the demo workflow. We don't claim all 22 live in one run.
preppedCloud Marketplace listing is prepped, not listed — Korea is outside the payment regions, which is exactly the A2A wedge.
zeroDesign partners: the program is open with 0 validated signups — pre-launch. We say so.
Social Seeding · Google for Startups AI Agents Challenge · Track 3 (Refactor) Business model Fixture proof Architecture Honest scope