代理协作

多 Agent 沙盒和工具

Status: active
Edit source

Each agent in a multi-agent setup can override the global sandbox and tool policy. This page covers per-agent configuration, precedence rules, and examples.


配置示例

Example 1: Personal + restricted family agent
json
{  "agents": {    "list": [      {        "id": "main",        "default": true,        "name": "Personal Assistant",        "workspace": "~/.openclaw/workspace",        "sandbox": { "mode": "off" }      },      {        "id": "family",        "name": "Family Bot",        "workspace": "~/.openclaw/workspace-family",        "sandbox": {          "mode": "all",          "scope": "agent"        },        "tools": {          "allow": ["read", "message"],          "deny": ["exec", "write", "edit", "apply_patch", "process", "browser"],          "message": {            "crossContext": {              "allowWithinProvider": false,              "allowAcrossProviders": false            }          }        }      }    ]  },  "bindings": [    {      "agentId": "family",      "match": {        "provider": "whatsapp",        "accountId": "*",        "peer": {          "kind": "group",          "id": "120363424282127706@g.us"        }      }    }  ]}

结果:

  • main agent:在主机上运行,拥有完整工具访问权限。
  • family agent:在 Docker 中运行(每个 agent 一个容器),仅允许 read 和当前对话消息发送。
Example 2: Work agent with shared sandbox
json
{  "agents": {    "list": [      {        "id": "personal",        "workspace": "~/.openclaw/workspace-personal",        "sandbox": { "mode": "off" }      },      {        "id": "work",        "workspace": "~/.openclaw/workspace-work",        "sandbox": {          "mode": "all",          "scope": "shared",          "workspaceRoot": "/tmp/work-sandboxes"        },        "tools": {          "allow": ["read", "write", "apply_patch", "exec"],          "deny": ["browser", "gateway", "discord"]        }      }    ]  }}
Example 2b: Global coding profile + messaging-only agent
json
{  "tools": { "profile": "coding" },  "agents": {    "list": [      {        "id": "support",        "tools": { "profile": "messaging", "allow": ["slack"] }      }    ]  }}

结果:

  • 默认 agent 获得编码工具。
  • support agent 仅支持消息传递(+ Slack 工具)。
Example 3: Different sandbox modes per agent
json
{  "agents": {    "defaults": {      "sandbox": {        "mode": "non-main",        "scope": "session"      }    },    "list": [      {        "id": "main",        "workspace": "~/.openclaw/workspace",        "sandbox": {          "mode": "off"        }      },      {        "id": "public",        "workspace": "~/.openclaw/workspace-public",        "sandbox": {          "mode": "all",          "scope": "agent"        },        "tools": {          "allow": ["read"],          "deny": ["exec", "write", "edit", "apply_patch"]        }      }    ]  }}

配置优先级

当同时存在全局配置(agents.defaults.*)和 agent 专属配置(agents.list[].*)时:

沙箱配置

agent 专属设置会覆盖全局设置:

Code
agents.list[].sandbox.mode > agents.defaults.sandbox.modeagents.list[].sandbox.scope > agents.defaults.sandbox.scopeagents.list[].sandbox.workspaceRoot > agents.defaults.sandbox.workspaceRootagents.list[].sandbox.workspaceAccess > agents.defaults.sandbox.workspaceAccessagents.list[].sandbox.docker.* > agents.defaults.sandbox.docker.*agents.list[].sandbox.browser.* > agents.defaults.sandbox.browser.*agents.list[].sandbox.prune.* > agents.defaults.sandbox.prune.*

工具限制

过滤顺序如下:

  • Tool profile

    tools.profileagents.list[].tools.profile

  • Provider tool profile

    tools.byProvider[provider].profileagents.list[].tools.byProvider[provider].profile

  • Global tool policy

    tools.allow / tools.deny

  • Provider tool policy

    tools.byProvider[provider].allow/deny

  • Agent-specific tool policy

    agents.list[].tools.allow/deny

  • Agent provider policy

    agents.list[].tools.byProvider[provider].allow/deny

  • Sandbox tool policy

    tools.sandbox.toolsagents.list[].tools.sandbox.tools

  • Subagent tool policy

    tools.subagents.tools,如适用。

  • Precedence rules
    • 每一层都可以进一步限制工具,但不能把前面层级已拒绝的工具重新授予回来。
    • 如果设置了 agents.list[].tools.sandbox.tools,它会替换该 agent 的 tools.sandbox.tools
    • 如果设置了 agents.list[].tools.profile,它会覆盖该 agent 的 tools.profile
    • 提供商工具键既可以接受 provider(例如 google-antigravity),也可以接受 provider/model(例如 openai/gpt-5.4)。
    Empty allowlist behavior

    如果该链路中的任何显式允许列表让本次运行没有任何可调用工具,OpenClaw 会在把 prompt 提交给模型之前停止。这是有意设计的:配置了缺失工具的 agent,例如 agents.list[].tools.allow: ["query_db"],应当在注册 query_db 的插件启用之前明确失败,而不是作为纯文本 agent 继续运行。

    工具策略支持 group:* 简写,它们会展开为多个工具。完整列表见 工具组

    每个 agent 的 elevated 覆盖项(agents.list[].tools.elevated)可以进一步限制特定 agent 的 elevated exec。详情见 Elevated mode


    从单 Agent 迁移

    迁移前(单 Agent)

    json
    {  "agents": {    "defaults": {      "workspace": "~/.openclaw/workspace",      "sandbox": {        "mode": "non-main"      }    }  },  "tools": {    "sandbox": {      "tools": {        "allow": ["read", "write", "apply_patch", "exec"],        "deny": []      }    }  }}

    迁移后(多 Agent)

    json
    {  "agents": {    "list": [      {        "id": "main",        "default": true,        "workspace": "~/.openclaw/workspace",        "sandbox": { "mode": "off" }      }    ]  }}

    工具限制示例

    只读 Agent

    json
    {  "tools": {    "allow": ["read"],    "deny": ["exec", "write", "edit", "apply_patch", "process"]  }}

    启用 Shell 执行但禁用文件系统工具

    json
    {  "tools": {    "allow": ["read", "exec", "process"],    "deny": ["write", "edit", "apply_patch", "browser", "gateway"]  }}

    仅通信

    json
    {  "tools": {    "sessions": { "visibility": "tree" },    "allow": ["sessions_list", "sessions_send", "sessions_history", "session_status"],    "deny": ["exec", "write", "edit", "apply_patch", "read", "browser"]  }}

    此配置文件中的 sessions_history 仍会返回一个有界且已清理的回忆视图,而不是原始转录转储。助手回忆会在编辑/截断之前剥离思考标签、<relevant-memories> 脚手架、纯文本工具调用 XML 负载(包括 <tool_call>...</tool_call><function_call>...</function_call><tool_calls>...</tool_calls><function_calls>...</function_calls> 和被截断的工具调用块)、降级的工具调用脚手架、泄漏的 ASCII/全角模型控制令牌,以及格式错误的 MiniMax 工具调用 XML。


    常见陷阱:"non-main"


    测试

    配置多 Agent 沙箱和工具后:

  • 检查 Agent 解析

    bash
    openclaw agents list --bindings
  • 验证沙箱容器

    bash
    docker ps --filter "name=openclaw-sbx-"
  • 测试工具限制

    • 发送一条需要受限工具的消息。
    • 验证 Agent 无法使用被拒绝的工具。
  • 监控日志

    bash
    tail -f "${OPENCLAW_STATE_DIR:-$HOME/.openclaw}/logs/gateway.log" | grep -E "routing|sandbox|tools"

  • 故障排除

    尽管设置了 `mode: 'all'`,Agent 仍未被沙箱隔离
    • 检查是否存在覆盖它的全局 agents.defaults.sandbox.mode
    • Agent 专用配置优先级更高,因此请设置 agents.list[].sandbox.mode: "all"
    尽管存在拒绝列表,工具仍然可用
    • 检查工具过滤顺序:全局 → Agent → 沙箱 → 子 Agent。
    • 每一层只能进一步限制,不能重新授予权限。
    • 使用日志验证:[tools] filtering tools for agent:${agentId}
    容器未按 Agent 隔离
    • 在 Agent 专用沙箱配置中设置 scope: "agent"
    • 默认值是 "session",即每个会话创建一个容器。

    相关

    Was this useful?