AgentForge: Autonomous Solution Architect for Google Cloud

Inspiration

Building production-grade autonomous agent systems is notoriously difficult. Developers face friction across four key areas:

1. Translating messy, real-world friction into structured multi-tool workflows.
2. Managing cross-session state without leaking tokens or paying heavy compute bills.
3. Implementing enterprise-grade security (Zero-Trust IAM, Model Armor prompt-injection defenses, OpenTelemetry distributed tracing).
4. Assembling reproducible spin-up environments under hackathon time constraints.

We built AgentForge to eliminate this friction entirely. It acts as an autonomous co-pilot and live testbed built natively for Google Cloud’s agent ecosystem.

What it does

AgentForge combines interactive agent playgrounds with an end-to-end Solution Architect & Devpost submission synthesizer:

• Track 1 (The Taskmaster Engine): Intercepts raw asynchronous event signals (Cloud Pub/Sub, Cloud Monitoring alerts) and executes end-to-end multi-tool remediations without human prompting.
• Track 2 (The Collaborative Partner): Leverages a persistent Firestore Memory Bank and dynamic RAG to maintain multi-turn context, generate adaptive comprehension checkpoints, and mutate live documents in real-time.
• Track 3 (The Fortified Enterprise Fleet): Simulates GEAP (Gemini Enterprise Agent Platform) mesh governance—incorporating Model Armor prompt-injection filtering, Zero-Trust IAM tokens, and OpenTelemetry trace trees.
• Cost & Credit Optimizer: Dynamically computes token inference overhead and enforces the "Flash First" design pattern to keep continuous autonomous runs well under the $150 Google Cloud credit budget.

How we built it

• Reasoning Engine: Powered by Gemini 3.7 Flash for sub-second tool routing, structured schema extraction, and autonomous agent orchestration.
• Agent Framework: Designed with Google Agent Development Kit (ADK) and Google GenAI SDK (@google/genai).
• Cloud Infrastructure: Architected for Google Cloud Run (scale-to-zero serverless container runtime), Google Cloud Pub/Sub (event ingestion), and Google Cloud Firestore (Memory Bank state & session store).
• Observability & Security: Employs Model Armor for PII redaction/prompt-injection scoring and Cloud Trace / OpenTelemetry for span latency breakdown.

Challenges we ran into

• Minimizing Autonomous Latency: Ensuring multi-step agent tool invocation chains completed in under 2 seconds without looping or hallucinating parameter schemas.
• Balancing Context vs. Token Budget: Architecting the memory bank to index high-signal cross-session facts without blowing up prompt token limits.
• Stateful Document Mutations: Implementing real-time synchronization between the agent's multi-turn conversational stream and the living workspace artifact.

Accomplishments that we're proud of

• Designed a decoupled, $0 idle cost architecture utilizing Google Cloud Run scale-to-zero.
• Built real-time interactive sandboxes covering all three official hackathon tracks with live telemetry.
• Automated the generation of complete, reproducible test suites and judging scripts conforming strictly to the 4-minute maximum pitch limit.

What we learned

• The Google ADK pattern combined with Gemini 3.7 Flash offers unprecedented speed for multi-turn autonomous loops compared to legacy monolithic orchestrators.
• Enforcing Model Armor at the API gateway layer provides deterministic guardrails against indirect prompt injection in autonomous pipelines.

What's next for AgentForge

• Adding native support for Google Gemma 2B edge agents on local worker nodes.
• Integrating Google Lyria and Veo APIs for autonomous multimodal asset generation during solution synthesis.
• Releasing the AgentForge ADK template directly to the open-source community.

Built With

• cloud-trace
• firestore
• gemini-3.7-flash
• google-adk
• google-cloud-run
• google-genai-sdk
• opentelemetry
• pub/sub
• react
• tailwind-css
• typescript