Managed Agents is a managed agent runtime provided by Model Studio, designed for long-running tasks such as multi-step tool calls, code execution, and file processing. The platform manages session state, sandbox environments, and tool execution. Agents autonomously execute commands, read and write files, install dependencies, and process data within isolated sandboxes, with event history persisted on the server side.
Differences from Agent Applications
|
Dimension |
Agent Applications |
Managed Agents |
|
Execution mode |
Stateless calls; context maintained on the application side |
Server-side session state management with interrupt and resume support |
|
Execution environment |
Shared runtime |
Isolated sandbox in cloud containers |
|
Event model |
Response-level streaming output |
Session-level SSE event stream with persistent event history |
|
Typical use cases |
Q&A, conversations, lightweight tasks |
Multi-step tool calls, code execution, file processing, and other long-running tasks |
Core Concepts
|
Concept |
Description |
|
Agent |
A configuration combining a model, system prompt, tools, MCP services, and Skills. Once created, an agent is referenced by ID and can be reused across multiple sessions. |
|
Environment |
The sandbox configuration in which a session runs, hosted as a cloud container managed by Model Studio. |
|
Session |
A running instance of an agent within a specified environment that executes tasks and generates output. |
|
Event |
A message exchanged between the application and the agent, including user messages, tool call results, and status changes. |
Supported Tools
Agents interact with the execution environment through the following tools:
-
Command execution: Run shell commands in the sandbox
-
File operations: Upload files to mount in the sandbox, or download from URLs; supports read, write, edit, find, and search
-
MCP services: Connect to external tool services
-
Skills: Mount pre-built tool combinations that encapsulate end-to-end task workflows
Workflow
-
Create an agent: Configure the model, system prompt, tools, and MCP services. Reference the agent by ID after creation.
-
Configure the environment: Create a cloud container sandbox and configure pre-installed packages as needed.
-
Start a session: Create a session instance that references the agent and environment configuration.
-
Send events and receive responses: Send user messages. The agent autonomously executes tool calls and pushes results through an SSE event stream. Event history is persisted on the server side.
-
Intervene or interrupt: Send new events during execution to guide the direction, or interrupt the current task.
Use Cases
-
Long-running tasks: Tasks involving multi-step tool calls that run for minutes to hours.
-
Code execution: Run scripts, install dependencies, and debug programs in a sandbox.
-
File processing: Read, write, search, and edit files within the sandbox.
-
Stateful sessions: Maintain context and file system state across multiple interaction turns.
-
Rapid development / low operational overhead: Focus on agent logic rather than building and maintaining agent loops, sandbox orchestration, or tool execution infrastructure from scratch.