DEV Community

Cover image for How We Built Crewmate: A 15-Agent Autonomous Enterprise Fleet for Content Creators
Lovejeet Singh
Lovejeet Singh

Posted on

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

#googlecloud #ai #agents #python

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.

Crewmate Logo

The 7-Layer GEAP Architecture

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

  1. Layer 1 — Presentation: React 19 SPA with 3D claymorphism UI, real-time multi-agent telemetry stream, and Web Speech API voice command gateway.
  2. 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).
  3. Layer 3 — Security & Identity: Model Armor middleware screening 15+ prompt injection patterns and redacting PII, combined with scoped RBAC identity matrices for each agent.
  4. 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.
  5. 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.
  6. Layer 6 — State & Memory Bank: Google Cloud Firestore (Native Mode) persistent store for creator preferences, minimum baseline rates, and historical negotiation logs.
  7. 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

Built by Lovejeet Singh & Sachit Babbar for the All Things Agentic Hackathon 2026.

Top comments (0)