Lead Agent · orchestrator
gemini-2.x on Vertex · deployed on Agent Engine
Parsed brief → "GlowSerum launch / $2k / US 22–34 skincare". Built plan, armed guardrails, fanned out to 5 sub-agents.
→ A2A: dispatch(plan) to Strategist
Strategist Agent
audience modeling · GMV target
3 hook angles ranked; predicted blended CPV $0.0098; GMV target $8–11k @ $2k spend.
mcp.trends.query(niche="skincare", geo="US")200 ok
MCP server: mock-tiktok-trends
→ A2A: handoff(audience_profile + angles) to Sourcing
Sourcing Agent
creator graph traversal
mcp.creators.search(tier="micro", geo="US", topic="skincare")200 · 1.4k hits
mcp.creators.score(ids=[…], model="cpv_predict")200 ok
MCP server: mock-tiktok-graph
3 tiers built. Paused for human approval → user picked mid-tier (18 creators).
→ A2A: handoff(shortlist_18) to Content
Content Agent
brief + caption generation
18 briefs generated; ran disclosure_check on each — all PASS (#ad + branded-content label enforced).
guardrail.disclosure_check(briefs=18)18/18 pass
→ A2A: handoff(approved_briefs) to Publisher
Publisher Agent
scheduling via mock Shop API
mcp.tiktok_shop.schedule(post, when=…) ×181 × 429 rate-limit
retry post#12 after backoff201 created
MCP server: mock-tiktok-shop
→ A2A: handoff(live_post_ids) to Analytics
Analytics Agent · running
verified-view metering · reallocation loop
mcp.metrics_sim.poll(post_ids)streaming…
guardrail.fraud_filter(raw=171,940)→ 146,321 verified
MCP server: metrics-simulator
Billable spend $1,463.21 / $2,000. Reallocating: pause 4, scale 3. Watching cap.
→ A2A: report(perf_snapshot) to Strategist · ⟳ loop
Guardrail · hard_budget_cap
armed — fires at $2,000.00 billable
⏸ will halt Publisher + Analytics spend, then emit ROI wrap-up to chat.