OpenClaw 多代理协同工作模式配置指南 - AI创想

# Lobster
Lobster is a workflow shell that lets OpenClaw run multi-step tool sequences as a single, deterministic operation with explicit approval checkpoints.
## Hook
Your assistant can build the tools that manage itself. Ask for a workflow,and 30 minutes later you have a CLI plus pipelines that run as one call. Lobster is the missing piece: deterministic pipelines, explicit approvals, and resumable state.
## Why
Today, complex workflows require many back-and-forth tool calls. Each call costs tokens,and the LLM has to orchestrate every step. Lobster moves that orchestration into a typed runtime:-**Onecall instead of many**: OpenClaw runs one Lobster tool call and gets a structured result.
-**Approvalsbuilt in**: Side effects (send email, post comment) halt the workflow until explicitly approved.
-**Resumable**: Halted workflows return a token; approve and resume without re-running everything.

复制代码

{broadcast:{strategy:"parallel","+15555550123":["developer","tester","security-auditor"]}}

复制代码

An **agent** is a fully scoped brain with its own:
- **Workspace** (files, AGENTS.md/SOUL.md/USER.md, local notes, persona rules).
- **State directory** (`agentDir`) for auth profiles, model registry, and per-agent config.
- **Session store** (chat history + routing state) under `~/.openclaw/agents/<agentId>/sessions`.
Auth profiles are **per-agent**. Each agent reads from its own:

复制代码

Main agent credentials are **not** shared automatically. Never reuse `agentDir`
across agents (it causes auth/session collisions). If you want to share creds,
copy `auth-profiles.json` into the other agent's `agentDir`.
Skills are per-agent via each workspace’s `skills/` folder, with shared skills
available from `~/.openclaw/skills`. See [Skills: per-agent vs shared](/tools/skills#per-agent-vs-shared-skills).

复制代码

`~/.openclaw/openclaw.json` (JSON5):
```js
{
agents: {
list: [
{
id: "home",
default: true,
name: "Home",
workspace: "~/.openclaw/workspace-home",
agentDir: "~/.openclaw/agents/home/agent",
},
{
id: "work",
name: "Work",
workspace: "~/.openclaw/workspace-work",
agentDir: "~/.openclaw/agents/work/agent",
},
],
},
// Deterministic routing: first match wins (most-specific first).
bindings: [
{ agentId: "home", match: { channel: "whatsapp", accountId: "personal" } },
{ agentId: "work", match: { channel: "whatsapp", accountId: "biz" } },
// Optional per-peer override (example: send a specific group to work agent).
{
agentId: "work",
match: {
channel: "whatsapp",
accountId: "personal",
peer: { kind: "group", id: "1203630...@g.us" },
},
},
],
// Off by default: agent-to-agent messaging must be explicitly enabled + allowlisted.
tools: {
agentToAgent: {
enabled: false,
allow: ["home", "work"],
},
},
channels: {
whatsapp: {
accounts: {
personal: {
// Optional override. Default: ~/.openclaw/credentials/whatsapp/personal
// authDir: "~/.openclaw/credentials/whatsapp/personal",
},
biz: {
// Optional override. Default: ~/.openclaw/credentials/whatsapp/biz
// authDir: "~/.openclaw/credentials/whatsapp/biz",
},
},
},
},
}

复制代码

Starting with v2026.1.6, each agent can have its own sandbox and tool restrictions:
```js
{
agents: {
list: [
{
id: "personal",
workspace: "~/.openclaw/workspace-personal",
sandbox: {
mode: "off", // No sandbox for personal agent
},
// No tool restrictions - all tools available
},
{
id: "family",
workspace: "~/.openclaw/workspace-family",
sandbox: {
mode: "all", // Always sandboxed
scope: "agent", // One container per agent
docker: {
// Optional one-time setup after container creation
setupCommand: "apt-get update && apt-get install -y git curl",
},
},
tools: {
allow: ["read"], // Only read tool
deny: ["exec", "write", "edit", "apply_patch"], // Deny others
},
},
],
},
}
Note: `setupCommand` lives under `sandbox.docker` and runs once on container creation.
Per-agent `sandbox.docker.*` overrides are ignored when the resolved scope is `"shared"`.
**Benefits:**
- **Security isolation**: Restrict tools for untrusted agents
- **Resource control**: Sandbox specific agents while keeping others on host
- **Flexible policies**: Different permissions per agent
Note: `tools.elevated` is **global** and sender-based; it is not configurable per agent.
If you need per-agent boundaries, use `agents.list[].tools` to deny `exec`.
For group targeting, use `agents.list[].groupChat.mentionPatterns` so @mentions map cleanly to the intended agent.
See [Multi-Agent Sandbox & Tools](/tools/multi-agent-sandbox-tools) for detailed examples.

复制代码

Fetch transcript for one session.
Parameters:
- `sessionKey` (required; accepts session key or `sessionId` from `sessions_list`)
- `limit?: number` max messages (server clamps)
- `includeTools?: boolean` (default false)
Behavior:
- `includeTools=false` filters `role: "toolResult"` messages.
- Returns messages array in the raw transcript format.
- When given a `sessionId`, OpenClaw resolves it to the corresponding session key (missing ids error).

复制代码

Send a message into another session.
Parameters:
- `sessionKey` (required; accepts session key or `sessionId` from `sessions_list`)
- `message` (required)
- `timeoutSeconds?: number` (default >0; 0 = fire-and-forget)
Behavior:
- `timeoutSeconds = 0`: enqueue and return `{ runId, status: "accepted" }`.
- `timeoutSeconds > 0`: wait up to N seconds for completion, then return `{ runId, status: "ok", reply }`.
- If wait times out: `{ runId, status: "timeout", error }`. Run continues; call `sessions_history` later.
- If the run fails: `{ runId, status: "error", error }`.
- Announce delivery runs after the primary run completes and is best-effort; `status: "ok"` does not guarantee the announce was delivered.
- Waits via gateway `agent.wait` (server-side) so reconnects don't drop the wait.
- Agent-to-agent message context is injected for the primary run.
- After the primary run completes, OpenClaw runs a **reply-back loop**:
- Round 2+ alternates between requester and target agents.
- Reply exactly `REPLY_SKIP` to stop the ping‑pong.
- Max turns is `session.agentToAgent.maxPingPongTurns` (0–5, default 5).
- Once the loop ends, OpenClaw runs the **agent‑to‑agent announce step** (target agent only):
- Reply exactly `ANNOUNCE_SKIP` to stay silent.
- Any other reply is sent to the target channel.
- Announce step includes the original request + round‑1 reply + latest ping‑pong reply.

复制代码

Behavior:
- Starts a new `agent:<agentId>:subagent:<uuid>` session with `deliver: false`.
- Sub-agents default to the full tool set **minus session tools** (configurable via `tools.subagents.tools`).
- Sub-agents are not allowed to call `sessions_spawn` (no sub-agent → sub-agent spawning).
- Always non-blocking: returns `{ status: "accepted", runId, childSessionKey }` immediately.
- After completion, OpenClaw runs a sub-agent **announce step** and posts the result to the requester chat channel.
- Reply exactly `ANNOUNCE_SKIP` during the announce step to stay silent.
- Announce replies are normalized to `Status`/`Result`/`Notes`; `Status` comes from runtime outcome (not model text).
- Sub-agent sessions are auto-archived after `agents.defaults.subagents.archiveAfterMinutes` (default: 60).
- Announce replies include a stats line (runtime, tokens, sessionKey/sessionId, transcript path, and optional cost).

复制代码

# Sub-agents
Sub-agents are background agent runs spawned from an existing agent run. They run in their own session (`agent:<agentId>:subagent:<uuid>`) and, when finished, **announce** their result back to the requester chat channel.
## Slash command
Use `/subagents` to inspect or control sub-agent runs for the **current session**:
- `/subagents list`
- `/subagents stop <id|#|all>`
- `/subagents log <id|#> [limit] [tools]`
- `/subagents info <id|#>`
- `/subagents send <id|#> <message>`
`/subagents info` shows run metadata (status, timestamps, session id, transcript path, cleanup).
Primary goals:
- Parallelize “research / long task / slow tool” work without blocking the main run.
- Keep sub-agents isolated by default (session separation + optional sandboxing).
- Keep the tool surface hard to misuse: sub-agents do **not** get session tools by default.
- Avoid nested fan-out: sub-agents cannot spawn sub-agents.
Cost note: each sub-agent has its **own** context and token usage. For heavy or repetitive
tasks, set a cheaper model for sub-agents and keep your main agent on a higher-quality model.
You can configure this via `agents.defaults.subagents.model` or per-agent overrides.

复制代码

## Tool
Use `sessions_spawn`:
- Starts a sub-agent run (`deliver: false`, global lane: `subagent`)
- Then runs an announce step and posts the announce reply to the requester chat channel
- Default model: inherits the caller unless you set `agents.defaults.subagents.model` (or per-agent `agents.list[].subagents.model`); an explicit `sessions_spawn.model` still wins.
- Default thinking: inherits the caller unless you set `agents.defaults.subagents.thinking` (or per-agent `agents.list[].subagents.thinking`); an explicit `sessions_spawn.thinking` still wins.
Tool params:
- `task` (required)
- `label?` (optional)
- `agentId?` (optional; spawn under another agent id if allowed)
- `model?` (optional; overrides the sub-agent model; invalid values are skipped and the sub-agent runs on the default model with a warning in the tool result)
- `thinking?` (optional; overrides thinking level for the sub-agent run)
- `runTimeoutSeconds?` (default `0`; when set, the sub-agent run is aborted after N seconds)
- `cleanup?` (`delete|keep`, default `keep`)

复制代码

Allowlist:
- `agents.list[].subagents.allowAgents`: list of agent ids that can be targeted via `agentId` (`["*"]` to allow any). Default: only the requester agent.
Discovery:
- Use `agents_list` to see which agent ids are currently allowed for `sessions_spawn`.

复制代码

Auto-archive:
- Sub-agent sessions are automatically archived after `agents.defaults.subagents.archiveAfterMinutes` (default: 60).
- Archive uses `sessions.delete` and renames the transcript to `*.deleted.<timestamp>` (same folder).
- `cleanup: "delete"` archives immediately after announce (still keeps the transcript via rename).
- Auto-archive is best-effort; pending timers are lost if the gateway restarts.
- `runTimeoutSeconds` does **not** auto-archive; it only stops the run. The session remains until auto-archive.

复制代码

## Announce
Sub-agents report back via an announce step:
- The announce step runs inside the sub-agent session (not the requester session).
- If the sub-agent replies exactly `ANNOUNCE_SKIP`, nothing is posted.
- Otherwise the announce reply is posted to the requester chat channel via a follow-up `agent` call (`deliver=true`).
- Announce replies preserve thread/topic routing when available (Slack threads, Telegram topics, Matrix threads).
- Announce messages are normalized to a stable template:
- `Status:` derived from the run outcome (`success`, `error`, `timeout`, or `unknown`).
- `Result:` the summary content from the announce step (or `(not available)` if missing).
- `Notes:` error details and other useful context.
- `Status` is not inferred from model output; it comes from runtime outcome signals.

复制代码

## Overview
Broadcast Groups enable multiple agents to process and respond to the same message simultaneously. This allows you to create specialized agent teams that work together in a single WhatsApp group or DM — all using one phone number.
Current scope: **WhatsApp only** (web channel).
Broadcast groups are evaluated after channel allowlists and group activation rules. In WhatsApp groups, this means broadcasts happen when OpenClaw would normally reply (for example: on mention, depending on your group settings).
## Use Cases
### 1. Specialized Agent Teams
Deploy multiple agents with atomic, focused responsibilities:
Group: "Development Team"
Agents:
- CodeReviewer (reviews code snippets)
- DocumentationBot (generates docs)
- SecurityAuditor (checks for vulnerabilities)
- TestGenerator (suggests test cases)

复制代码

### Basic Setup
Add a top-level `broadcast` section (next to `bindings`). Keys are WhatsApp peer ids:
- group chats: group JID (e.g. `120363403215116621@g.us`)
- DMs: E.164 phone number (e.g. `+15551234567`)
```json
{
"broadcast": {
"120363403215116621@g.us": ["alfred", "baerbel", "assistant3"]
}
}
**Result:** When OpenClaw would reply in this chat, it will run all three agents.

复制代码

### Session Isolation
Each agent in a broadcast group maintains completely separate:
- **Session keys** (`agent:alfred:whatsapp:group:120363...` vs `agent:baerbel:whatsapp:group:120363...`)
- **Conversation history** (agent doesn't see other agents' messages)
- **Workspace** (separate sandboxes if configured)
- **Tool access** (different allow/deny lists)
- **Memory/context** (separate IDENTITY.md, SOUL.md, etc.)
- **Group context buffer** (recent group messages used for context) is shared per peer, so all broadcast agents see the same context when triggered
This allows each agent to have:
- Different personalities
- Different tool access (e.g., read-only vs. read-write)
- Different models (e.g., opus vs. sonnet)
- Different skills installed
### Example: Isolated Sessions

复制代码

const SessionsSendToolSchema = Type.Object({
sessionKey: Type.Optional(Type.String()),
label: Type.Optional(Type.String({ minLength:1, maxLength:SESSION_LABEL_MAX_LENGTH})),
agentId: Type.Optional(Type.String({ minLength:1, maxLength:64})),
message: Type.String(),
timeoutSeconds: Type.Optional(Type.Number({ minimum:0})),});

复制代码

# Lobster
Lobster is a workflow shell that lets OpenClaw run multi-step tool sequences as a single, deterministic operation with explicit approval checkpoints.
## Hook
Your assistant can build the tools that manage itself. Ask for a workflow, and 30 minutes later you have a CLI plus pipelines that run as one call. Lobster is the missing piece: deterministic pipelines, explicit approvals, and resumable state.
## Why
Today, complex workflows require many back-and-forth tool calls. Each call costs tokens, and the LLM has to orchestrate every step. Lobster moves that orchestration into a typed runtime:
- **One call instead of many**: OpenClaw runs one Lobster tool call and gets a structured result.
- **Approvals built in**: Side effects (send email, post comment) halt the workflow until explicitly approved.
- **Resumable**: Halted workflows return a token; approve and resume without re-running everything.

复制代码