Multi-Cloud Deployment¶

Overview¶

ADEPT is designed for multi-cloud operation from the ground up. OCI-compliant container images, provider-neutral Helm charts, and pluggable LLM routing allow a single ADEPT deployment to span AWS, Azure, and GCP -- or federate multiple deployments across clouds using the A2A gateway mesh.

Multi-Cloud Strategy¶

The architecture separates concerns into three layers:

graph TD
    subgraph Application["Application Layer (Portable)"]
        HELM[Helm Chart]
        DOCKER[OCI Images]
    end
    subgraph IaC["Infrastructure Layer (Provider-Specific)"]
        CDK[AWS CDK]
        PULUMI[Azure Pulumi]
        TF[GCP Terraform]
    end
    subgraph LLM["LLM Routing Layer (Provider-Neutral)"]
        LITELLM[LiteLLM Router]
    end
    HELM --> CDK
    HELM --> PULUMI
    HELM --> TF
    Application --> LLM

Provider Matrix¶

LLM Provider Configuration¶

ADEPT routes LLM requests through LiteLLM, supporting multiple providers simultaneously. The model name prefix determines the provider:

# Azure OpenAI
DEFAULT_LLM_MODEL=azure/gpt-4o

# AWS Bedrock
DEFAULT_LLM_MODEL=bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0

# Anthropic Direct
DEFAULT_LLM_MODEL=anthropic/claude-sonnet-4-5-20250929

# Google Vertex AI
DEFAULT_LLM_MODEL=vertex_ai/gemini-2.0-flash

# Local Ollama
DEFAULT_LLM_MODEL=ollama/llama3.1

# NVIDIA NIMs
DEFAULT_LLM_MODEL=nvidia_nim/meta/llama-3.1-70b-instruct

Cross-Cloud Federation¶

The A2A (Agent-to-Agent) protocol enables ADEPT instances in different clouds to communicate:

graph LR
    subgraph AWS
        GW1[Gateway A]
    end
    subgraph Azure
        GW2[Gateway B]
    end
    subgraph GCP
        GW3[Gateway C]
    end
    REG[Gateway Registry]
    GW1 <--> REG
    GW2 <--> REG
    GW3 <--> REG
    GW1 <-->|A2A Protocol| GW2
    GW2 <-->|A2A Protocol| GW3

Each gateway registers with a shared Gateway Registry. Agents on one gateway can discover and invoke tools hosted on any federated peer, enabling cross-cloud workflows without data migration.

Deployment Patterns¶

Pattern 1: Single-Cloud Full Stack¶

Deploy all components in one cloud provider. Simplest operational model.

# AWS example
cd infra/aws/cdk && cdk deploy --all
helm install adept infra/helm/agentic-framework/ -f values-aws.yaml

Pattern 2: Multi-Cloud Federation¶

Deploy independent stacks per cloud, federate via A2A:

1. Deploy Stack A in AWS (owns HPC tools)
2. Deploy Stack B in Azure (owns Azure OpenAI models)
3. Register both gateways in shared registry
4. Agents in either stack can invoke tools from both

Pattern 3: Hybrid Cloud + On-Premises¶

Combine cloud deployments with on-premises Docker Compose instances for air-gapped or data-sovereignty requirements. The A2A mesh treats all gateways identically regardless of hosting.