agent-management

name	agent-management
description	Create, list, and configure OrgOps agents from native agents. Use when the user asks an agent to create agents, list agents, set up WRAPPED agents, OpenClaw wrappers, sidecars, or Cursor coding harnesses.

Agent Management

Use the native runner tools for agent operations:

• agents_search: list or filter existing agents.
• agents_create: create CLASSIC, RLM_REPL, or WRAPPED agents.
• agents_update: patch an existing agent's config, lifecycle state, skills, model, workspace, or wrapped runtime config.
• events_channel_participant_add or events_channel_join: add an agent to a channel when it should receive channel events.

List Agents

Call agents_search first when you need to inspect existing agents, avoid duplicate names, or find runner assignments.

Example:

{
  "nameContains": "coordinator",
  "limit": 20
}

Create Native Agents

For a normal OrgOps-managed LLM agent, use mode: "CLASSIC" or omit mode. Pick a concrete modelId, set a workspace, and enable any required skills.

Example:

{
  "name": "ResearchAgent",
  "mode": "CLASSIC",
  "modelId": "openai:gpt-4o-mini",
  "workspacePath": ".orgops-data/workspaces/ResearchAgent",
  "enabledSkills": ["tavily"],
  "alwaysPreloadedSkills": ["tavily"],
  "joinCurrentChannel": true
}

Create Wrapped Agents

For a wrapped agent, set:

• mode: "WRAPPED"
• modelId: "wrapped:none" if you provide it explicitly
• memoryContextMode: "OFF" or omit it; OrgOps forces wrapped agents to memory OFF
• enabledSkills: [] and alwaysPreloadedSkills: []
• a wrappedConfig object with kind, harness, optional setup, optional sidecars, and runtime

Wrapped agents do not use OrgOps model calls, prompts, memory, skills, or allowOutsideWorkspace for turns. The external runtime owns its session, tools, memory, prompt behavior, and filesystem policy through its own settings.

Update Existing Agents

Use agents_update instead of direct database writes. For wrapped agents, prefer wrappedConfigJson when patching a whole config from generated JSON; it must parse to a JSON object.

Example:

{
  "agentName": "OrgOpsCoordinator",
  "wrappedConfigJson": "{\"kind\":\"openclaw\",\"harness\":\"command\",\"runtime\":{\"command\":\"./run-openclaw.sh\",\"cwd\":\".orgops-data/workspaces/OrgOpsCoordinator/wrapper\",\"parse\":\"json-payloads\",\"timeoutMs\":600000}}",
  "desiredState": "RUNNING"
}

OpenClaw With Gateway Sidecar

Use OpenClaw when the wrapped runtime is an OpenClaw agent. The OpenClaw sidecar should be the OpenClaw Gateway; OrgOps starts it before turns and the runtime command sends each turn through that gateway.

Example agents_create arguments:

{
  "name": "OrgOpsCoordinator",
  "mode": "WRAPPED",
  "workspacePath": ".orgops-data/workspaces/OrgOpsCoordinator",
  "wrappedConfig": {
    "kind": "openclaw",
    "harness": "command",
    "session": { "scope": "per-channel" },
    "setup": {
      "command": "npm install",
      "cwd": ".orgops-data/workspaces/OrgOpsCoordinator/wrapper",
      "timeoutMs": 120000
    },
    "sidecars": [
      {
        "name": "gateway",
        "command": "npx openclaw gateway --force",
        "cwd": ".orgops-data/workspaces/OrgOpsCoordinator/wrapper",
        "restart": true,
        "restartDelayMs": 2000,
        "timeoutMs": 0
      }
    ],
    "runtime": {
      "command": "npx openclaw agent --agent coordinator --session-id \"$ORGOPS_WRAPPED_SESSION_ID\" --message \"$ORGOPS_WRAPPED_MESSAGE\" --json",
      "cwd": ".orgops-data/workspaces/OrgOpsCoordinator/wrapper",
      "parse": "json-payloads",
      "timeoutMs": 600000
    }
  },
  "joinCurrentChannel": true
}

OpenClaw setup notes:

• Store wrapper source in the agent workspace or configure source if it comes from a repo.
• Put secrets in OrgOps package secrets, not in wrappedConfig; setup, sidecars, and runtime receive package secrets as env.
• Use ORGOPS_WRAPPED_SESSION_ID to keep OpenClaw state stable per channel or per agent.
• Configure OpenClaw's gateway, agents, model allowlist, channel bridges, sandbox, and filesystem policy in OpenClaw itself.
• Keep the gateway sidecar idempotent and restartable.

Cursor Coding Harness

Use Cursor wrapped mode when the agent should delegate coding work to Cursor tooling rather than the OrgOps LLM loop. Treat it as a high-power coding harness: give it an isolated workspace, a narrow prompt/message bridge, and a long turn timeout.

Example agents_create arguments:

{
  "name": "CursorCoder",
  "mode": "WRAPPED",
  "workspacePath": ".orgops-data/workspaces/CursorCoder",
  "wrappedConfig": {
    "kind": "cursor",
    "harness": "command",
    "session": { "scope": "per-agent" },
    "runtime": {
      "command": "cursor-agent -p \"$ORGOPS_WRAPPED_MESSAGE\" --output-format text",
      "cwd": ".",
      "parse": "text",
      "timeoutMs": 1800000
    }
  }
}

Cursor setup notes:

• Configure CURSOR_API_KEY as an OrgOps secret in package cursor.
• Configure Cursor's filesystem and permission policy in the Cursor harness settings; OrgOps allowOutsideWorkspace does not apply to wrapped agents.
• Prefer session.scope: "per-agent" for coding continuity across channels; use per-channel when channel isolation matters.
• Ask before creating or modifying repositories outside the configured workspace.