How We Built Crewmate: A 15-Agent Autonomous Enterprise Fleet for Content Creators

Lovejeet Singh — Mon, 31 Aug 2026 21:03:44 +0000

Disclaimer: This article was created for the purposes of entering the **All Things Agentic Hackathon 2026* on Devpost under Track 3: The Fortified Enterprise Fleet.*

The Unlikely Hero Thesis

Enterprise agent fleets and governance frameworks are traditionally built for Fortune 500 CIOs. Goldman Sachs analysts have Model Armor; JPMorgan compliance officers have rate-limited agent gateways.

Meanwhile, over 50 million content creators worldwide manage multi-million-view media businesses with spreadsheets, manual checklists, and gut instinct. A single hidden 12-month exclusivity clause in a sponsorship contract can cost a creator $80,000+ in lost brand deals. A single FTC disclosure oversight can lead to platform demonetization.

We asked a simple question: What if we gave every content creator the exact same enterprise-grade multi-agent fleet that Fortune 500 enterprises rely on?

That question became Crewmate.

The 7-Layer GEAP Architecture

Crewmate implements a strict 7-layer decoupled architecture adhering 100% to the Gemini Enterprise Agent Platform (GEAP) specification:

Layer 1 — Presentation: React 19 SPA with 3D claymorphism UI, real-time multi-agent telemetry stream, and Web Speech API voice command gateway.
Layer 2 — API Gateway: FastAPI on Cloud Run with sliding-window rate limiting (120 req/min) and 3-state circuit breakers (CLOSED → OPEN → HALF_OPEN).
Layer 3 — Security & Identity: Model Armor middleware screening 15+ prompt injection patterns and redacting PII, combined with scoped RBAC identity matrices for each agent.
Layer 4 — Orchestration Engine: Fleet Orchestrator Supervisor (gemini-3.1-pro-preview) that decomposes high-level goals into parallel sub-tasks and queries the Firestore Agent Registry.
Layer 5 — Autonomous 15-Agent Fleet: 14 specialized worker agents built with Google ADK running on gemini-3.7-flash, Google Veo 3.1, Gemini 3 Pro Image, Lyria AI, and Gemma 2 9B.
Layer 6 — State & Memory Bank: Google Cloud Firestore (Native Mode) persistent store for creator preferences, minimum baseline rates, and historical negotiation logs.
Layer 7 — Observability & Telemetry: OpenTelemetry-compliant distributed span tracing capturing token counts, tool call latencies, and security audit logs.

The 15-Agent Fleet in Action

Our multi-agent system operates across 6 functional domain clusters:

Legal & Business Intelligence:
- Contract Reviewer: Scans PDF contracts in seconds, extracts clauses, flags predatory terms, and drafts lawyer-grade counter-proposals.
- Revenue Optimizer: Benchmarks deal CPMs against market rates and uncovers negotiation leverage.
- Brand Safety: Evaluates sponsor controversies and alignment with creator values.
Content & Compliance:
- Content Compliance: Audits video metadata against FTC 16 CFR § 255 and scans for copyrighted audio, suggesting Lyria AI royalty-free music replacements.
- Distribution Manager: Generates algorithmic YouTube SEO descriptions and Instagram Reels hashtags.
Multimodal Media Studio:
- AI Video Cinematographer: Synthesizes 8-second 1080p B-roll clips via Google Veo 3.1 with camera direction.
- Thumbnail Director: Directs Gemini 3 Pro Image to produce 1376×768 native widescreen thumbnails with predicted CTR scores up to 94%.
Growth & Community:
- Hook & Script Architect: Retention engineering for the first 3 seconds of video + beat-by-beat scripts.
- Trend Radar: Scans breakout viral topics and calculates velocity scores.
- Community Guardian: Uses Gemma 2 9B for edge sentiment clustering and toxic comment moderation.
- Threat Sentinel: Monitors the fleet itself for prompt injection anomalies.

Deployment & Production Engineering

The entire platform is production-deployed on Google Cloud Run in us-central1:

Single-Stage Container: We unified the FastAPI backend and pre-built React 19 static assets into a single Python 3.13-slim container, eliminating cross-service networking overhead.
Firestore Persistence: 5 collections (agents, memory, contracts, traces, armor_logs) handle state, memory, and telemetry with real-time UI synchronization.
Autonomous Execution: Uploading a PDF contract triggers 4 parallel agents (Reviewer, Revenue, Brand Safety, and Report Generator) with zero follow-up prompting needed.

Try It Out

Live Cloud Run App: https://crewmate-285381944529.us-central1.run.app
GitHub Repository: https://github.com/z-lovejeet/Crewmate
Architecture Blueprint: Available in the root repository as Crewmate_System_Architecture.pdf