⚙️ Atlas Magenta

Build Agents,
Not Infrastructure

The fully managed platform for deploying AI agents to production. MongoDB's unified database, model-neutral architecture, and enterprise-grade isolation — so you focus on your agent logic, not the plumbing.

Why Magenta → See It In Action

Why Magenta

Three capabilities that set this platform apart — and why enterprises building production agents choose MongoDB.

🍃

Unified Database Power

One cluster for all agent memory — conversations, knowledge embeddings, entity graphs, checkpoints, and audit logs. No sync pipelines. No consistency bugs.

Atlas Vector Search Voyage AI Embeddings $graphLookup Change Streams TTL Indexes Document Model
🔀

Model & Cloud Neutral

Bring any LLM via Model Proxy. Support for LangGraph, CrewAI, Google ADK, and custom frameworks. Designed to be truly multi-cloud.

Any LLM Provider Multi-Framework MCP Server Support Agent-to-Agent Multi-Cloud Roadmap
🏛️

Enterprise Scale & Security

5-layer tenant isolation (compute, data, network, secrets, control plane), CSFLE encryption, private networking, RBAC, and per-execution resource limits.

Tenant Isolation CSFLE Private Networking RBAC + OIDC OpenTelemetry Audit Logging
☁️ Fully Managed

Platform Features

Everything your agents need to run in production — orchestration, execution, memory, observability, and security. All managed by MongoDB. Click any component to explore.

Platform Components

Click any component to see how it works and what MongoDB handles for you.

🎯
Orchestrator
Routes calls, enforces policies
🤖
Agent Executor
Isolated agent container runtime
🔧
Tool Executor
Sandboxed tool execution with MCP
🧠
Memory Service
Short-term & long-term memory
🍃
Customer Data Store
Atlas cluster in customer's org
📊
Observability
Traces, logs, cost tracking
🔐
Secrets Store
Encrypted credentials via CSFLE
🎯 Orchestrator

Orchestrator

The central proxy that routes calls between agents, tools, LLMs, and other connections. It reviews and enforces authorization and policies established at the Organization and Project level.

Request Routing Policy Enforcement AuthZ Checks Checkpoint Relay Rate Limiting
MongoDB Role

Relays Executor Checkpoints to the Customer Data Store for durable execution. Execution Logs and Traces flow through the Orchestrator to the Runtime Data Store. All routing decisions are persisted for audit compliance.

Isolation Guarantees

✓ Inter-component traffic scoped to single Project
✓ Network policies deny cross-tenant traffic by default
✓ Per-execution limits on tool calls, LLM calls, and duration

See It In Action

Watch a real agent request flow through the platform — from API ingress through orchestration, execution, memory access, and observability. Step through or auto-run.

☁️ Fully Managed

Every component below — API Gateway, Orchestrator, Executors, Memory, and Observability — is hosted, secured, and scaled by MongoDB. You deploy your agent code. We run everything else.

See it for your industry:
🏗️
Insurance Agent — web-chat MANAGED
Project: insurance-dev · Workspace: web-chat · Your preferred region
🌐
API Gateway
Client → AuthN
→
🎯
Orchestrator
Route + Policy
→
🤖
Agent Exec
Your Framework
→
🔧
Tool Exec
MCP Server
→
🧠
Memory
Short + Long Term
→
🍃
Data Store
Atlas Cluster
→
📊
Observe
Traces + Cost
📡 Call Stack 0 calls
Select a scenario, then click Trigger to begin...
0
Hops
0
Tools
0
Reads
0
Writes
📋Execution Trace
OpenTelemetry
📋
Execution trace will build as the request flows through components
🤖 Agent Response
💭 Current Phase
Select an industry above to see how Magenta runs your agent end-to-end.
Pick an industry above to see a domain-specific agent in action.

Resource Hierarchy

Agents are organized in a three-level hierarchy mirroring Atlas. Click each level to see what it contains and how isolation boundaries work.

→ →
🏢

Organization

Top-level container and tenant isolation boundary. Data, credentials, and resources never cross Organization boundaries. Customers pay for the platform at this level. Supports federated identity providers and defines policies that cascade to all child Projects.

Tenant Isolation Billing Scope AuthN / Federation Image Registry Policies

Example: State Farm

resource-hierarchy.txt
Organization: State Farm
├── Project: Claims Agent Dev
│   ├── Workspace: Web Chat
│   └── Workspace: Phone
├── Project: Claims Agent Prod
└── Project: Underwriting Agent

Developer Workflow

From local development to production deployment — the complete agent engineering lifecycle using the Magenta SDK, CLI, and platform tools.

01
Define Agent
Magenta SDK + your framework
02
Develop Locally
magenta dev — local runtime emulation
03
Build & Package
magenta build — container images to registry
04
Deploy to Cloud
magenta deploy — multi-tenant cloud
05
Observe & Monitor
OpenTelemetry traces + Agent Playground
📦 Magenta SDK

Define Agent

Define agents using the Magenta Python SDK with your framework of choice — LangGraph, CrewAI, Google ADK, or custom code. The SDK handles tool isolation, model access, memory, and durable execution via checkpointing.

agent.py
from magenta import Agent, Tool
from langgraph.graph import StateGraph

# Define tools with Magenta SDK
@Tool("lookup_policy")
def lookup_policy(policy_id: str):
    """Look up an insurance policy by ID."""
    return db.policies.find_one({"_id": policy_id})

# Build agent with your preferred framework
graph = StateGraph(AgentState)
graph.add_node("agent", agent_node)
graph.add_node("tools", tool_node)

agent = Agent(
    graph=graph.compile(),
    model="claude-sonnet",
    memory="project",  # uses project-level memory
)
⌨️

CLI

magenta dev, magenta build, magenta deploy — full lifecycle from terminal.

🏗️

Terraform

IaC with magenta_project, magenta_workspace resources for GitOps workflows.

🖥️

Web UI

Agent Playground, execution traces, cost monitoring, and policy management in the browser.

From Prototype to Production

The same platform that runs your first agent runs thousands. Here's what Magenta handles at every stage of your journey.

🧪

Prototype

1-3 agents · single developer

Go from idea to working agent in an afternoon. magenta dev emulates the full production runtime locally so you can iterate without deploying.

✓ Local runtime with production parity
✓ Agent Playground for interactive testing
✓ Single Project, single Workspace
✓ Atlas cluster for memory and data
✓ OpenTelemetry traces for debugging
🏗️

Production

10-50 agents · platform team

Enterprise isolation and governance from day one. Dedicated executors, private networking, and policy enforcement — all managed so your team focuses on agent logic.

✓ Isolated Agent and Tool Executor sandboxes
✓ Dev / Stage / Prod Projects
✓ Private networking to Atlas and endpoints
✓ RBAC + OIDC identity federation
✓ Per-execution resource limits and policies
✓ CDC pipelines for enterprise data ingestion
✓ Terraform and CLI for GitOps workflows
🌍

Scale

100+ agents · multi-team organization

The same platform that runs one agent runs thousands. Horizontal scaling, multi-agent coordination, and global deployment — without re-architecting.

✓ Horizontal sharding by agent_id
✓ Multi-agent coordination across Projects
✓ Agent discovery and enterprise catalog
✓ Per-Organization billing and usage metering
✓ Multi-region replication
✓ Custom memory consolidation workers
📈

Horizontal Sharding

Memory collections shard by agent_id. Atlas auto-balances chunks across nodes as agent count grows. No application-level sharding logic needed.

⚡

Change Streams

Real-time event propagation between agents without polling. CDC pipelines keep enterprise data from external sources synchronized into the Customer Data Store automatically.

🛡️

Tenant Isolation

Five-layer isolation: compute (per-execution sandboxes), data (scoped to Organization), network (deny cross-tenant by default), secrets (CSFLE), and control plane (RBAC). No misconfiguration can leak data.