Agents overview

The Agents service turns a folder of Python code into a secure, production-ready HTTPS endpoint. You write your agent logic, and Crusoe AI Platform handles the infrastructure, scaling, and operational management.

This guide introduces the core concepts behind Crusoe AI Platform Agents, supported frameworks, the deploy lifecycle, and how Crusoe compares to other cloud providers.

---

What the Agents service is

You write an AI agent using Python code that talks to a language model and executes tools. When you upload your project, the platform automatically packages your code into a container image, deploys it, and provides a public HTTPS URL.

Agents scale automatically based on traffic:

• Automatic scaling: The platform scales your instance count up as traffic increases.
• Scale-to-zero: When your agent receives no traffic, it scales down to zero instances so you only pay for active compute time. The next incoming request automatically wakes the agent up.

Deploying your first agent

You can deploy using the Console UI or the platform CLI.

Using the Console UI

1. Open the Console and navigate to Agents.
2. Click Create Agent.
3. Select your framework (ADK, LangGraph, or CrewAI) and upload your project directory or repository.
4. Click Deploy.

Using the CLI

Run the deploy command from your terminal:

platformctl deploy ./my-agent

---

The harness architecture

Every deployed agent runs inside a managed runtime environment called the harness. The harness is a web server built into every agent container image. It receives incoming HTTP requests, passes messages to your code, and manages state and background services.

The harness automatically provides these built-in capabilities:

• Managed Inference Model: Access models via the foundry_model() helper function. Models follow the OpenAI-compatible HTTP protocol, allowing standard SDKs and tools to work seamlessly. Swap models anytime using environment variables.
• Session Management: Each conversation tracks state using a session_id. The platform stores and replays message history automatically so the model retains context across turns. See Sessions.
• Long-Term Memory: Store persistent knowledge using the memorize action and query it using the built-in search_memory tool. See Memory.
• Isolated Code Sandbox: The run_python tool executes Python code in a secure, single-use sandbox separate from your primary agent container. See Tools.

Because the harness handles HTTP routing and protocols, every deployed agent exposes a standardized REST API, regardless of the underlying framework.

---

Supported frameworks

Crusoe AI Platform supports three popular open-source agent frameworks as first-class citizens. Select a framework during deployment based on your project structure:

Framework	Entry File	Required Export	Guide
ADK (Google Agent Development Kit)	agent.py	root_agent module object	ADK guide
LangGraph	graph.py	Compiled graph module object	LangGraph guide
CrewAI	crew.py	crew module object	CrewAI guide

The platform runs unmodified open-source ADK, LangGraph, or CrewAI code. Base container images include pre-installed helper packages (crusoe_adk, crusoe_langchain, crusoe_crewai) that configure models and built-in tools automatically.

---

The deploy lifecycle

Agent deployments move through distinct lifecycle states. You can monitor state transitions using platformctl status or the platform API (GET /v1/agents/{name}):

Lifecycle states

• building: The platform builds your container image and installs dependencies listed in requirements.txt.
• deploying: The container image build is complete. The platform creates a new immutable revision and prepares traffic routing.
• ready: The revision is active and ready to process requests. The ready status flag equals true.
• failed: Deployment encountered an error. Detailed error logs and output tails are published to the agent's message field. See Deploy guide and Troubleshooting.

---

Platform comparison

The following table compares Crusoe AI Platform Agents with managed agent solutions from other cloud providers:

Feature	Crusoe AI Platform Agents	Vertex AI Agent Engine (GCP)	Bedrock AgentCore (AWS)	Azure Foundry Agent Service
Deploy Path	Unified command: platformctl deploy or Console UI	Python SDK (agent_engines.create())	agentcore deploy CLI tool	azd deploy or Azure Portal
Supported Frameworks	Native ADK, LangGraph, CrewAI	ADK primary	Custom framework SDKs	Azure Agent Framework, LangGraph
Session & Memory Management	Built-in automatic harness state	Configured via separate cloud services	Requires independent component setup	Managed threads via Azure infrastructure
Scaling	Automatic scale-to-zero compute	Configured instance pools	Auto-scaled container tasks	Container app auto-scaling

Note: Multi-region replication, custom SLA guarantees, and enterprise compliance controls (such as HIPAA compliance) are currently unavailable.

For a full service-by-service mapping across cloud providers, see Service Mapping.

---

Next steps

• Deploy an Agent — Learn directory structures, deployment steps, and configuration options.
• Framework Guides — Build agents with ADK, LangGraph, or CrewAI.
• Invoke an Agent — Send requests, stream responses, and manage credentials.
• Sessions & Memory — Manage conversation context and long-term memory banks.
• Tools — Implement custom tools and leverage built-in code sandboxes.
• Secrets and Environment Variables — Manage credentials and runtime configurations.
• Embedded Chat Widget — Embed your agent as an interactive chat widget on any website.