invocation_patterns
How agents are triggered. Mechanisms include natural language detection, @-mention syntax, slash commands, CLI flags, or automatic delegation.
Type: object Content type: agents
Provider Support
Section titled “Provider Support”| Provider | Supported | Mechanism |
|---|---|---|
| claude-code | ✓ | natural language (Claude auto-delegates), @agent-<name> @-mention, or --agent flag at startup for session-wide agent; 3 patterns |
| codex | ✓ | spawn-agent tool at LLM discretion; LLM selects role from dynamically-built role descriptions combining built-in and user-defined roles |
| copilot-cli | ✓ | /agent slash command, natural language instruction, auto-inference (suppressible with disable-model-invocation), or --agent flag; 4 patterns |
| cursor | ✓ | Custom agents are invoked through the Cursor IDE agent picker; background-enabled agents can also run concurrently via the background agents UI. |
| factory-droid | ✓ | exposed as subagent_type targets for the Task tool; droids enabled by default with live progress reporting in UI |
| kiro | ✓ | keyboard shortcut binding for IDE invocation; welcomeMessage shown on agent start; manual selection via agent picker |
| opencode | ✓ | Agent name derives from the filename (configEntryNameFromPath). Agents with mode: primary can be selected as the session agent; agents with mode: subagent are invoked by other agents via description-driven matching; mode: all allows both. |
| roo-code | ✓ | Users switch modes via the mode selector or /mode command; each mode is a distinct operating persona with its own roleDefinition and tool groups |
| crush | ✗ | not documented |
| windsurf | ✗ | not documented |
| zed | ✗ | not documented |