Skip to main content
This is preview documentation. The official release is not yet available.
Managed Agents execute the platform Harness. Distinguish model connections, execution environments, tools and work entry points when extending an Agent.

​
Model integrations

The Java SDK provides model extensions for OpenAI and compatible APIs, Anthropic, Gemini and Ollama. These are integration options; their existence in the SDK does not make every provider part of the standard Service image. The selected model must support the tool calls and input types your task uses.

​
Tools and MCP

Harness and the chosen Environment provide built-in tools. agent_toolset controls built-in tools, while mcp_toolset references an mcpServerName. Define the connection in mcpServers. Common mcpServers fields are name, transport, url, command, args, headers, env, queryParams, enableTools, disableTools, required, initializationTimeout and timeout. Use command/args for stdio and url for a network endpoint. Timeout fields accept duration strings supported by the adapter, such as PT30S. enabled controls tool availability; permissionPolicy.type accepts always_allow, always_ask or deny. A configured tool still needs permission to execute; requests requiring confirmation enter the corresponding approval flow.

​
Skills and internal Subagents

Workspace holds AGENTS.md, Skill files and Subagent declarations. Skills provide repeatable procedures and supporting material. Subagents provide specialist delegation inside an Agent. Install system dependencies in the actual Environment. Use a Team for independent ownership, persistent discussion, mixed runtimes or a combined deliverable. Internal Subagents are not automatically exposed as independent Team services.

​
Knowledge, credentials and entry points

  • Memory Store: bound shared knowledge read on demand during execution.
  • Vault: tool connection credentials; it does not configure every model connection.
  • Chat: personal conversations. Issue: delivery and acceptance.
  • Automation, Workflow and Team: schedule, compose or delegate Managed work.
  • Endpoint and Channel: expose capabilities through published protocols and message routing.
Keep a minimal input, expected tool call and checkable output for each added capability. Verify the Agent alone before adding it to a Team or Workflow.