openclaw/docs/tools/subagents.md

summary, read_when, title, sidebarTitle

summary	read_when	title	sidebarTitle
Spawn isolated background agent runs that announce results back to the requester chat	You want background or parallel work via the agent You are changing sessions_spawn or sub-agent tool policy You are implementing or troubleshooting thread-bound subagent sessions	Sub-agents	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. Each sub-agent run is tracked as a background task.

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.
• Support configurable nesting depth for orchestrator patterns.

**Cost note:** each sub-agent has its own context and token usage by default. For heavy or repetitive tasks, set a cheaper model for sub-agents and keep your main agent on a higher-quality model. Configure via `agents.defaults.subagents.model` or per-agent overrides. When a child genuinely needs the requester's current transcript, the agent can request `context: "fork"` on that one spawn.

Slash command

Use /subagents to inspect or control sub-agent runs for the current session:

/subagents list
/subagents kill <id|#|all>
/subagents log <id|#> [limit] [tools]
/subagents info <id|#>
/subagents send <id|#> <message>
/subagents steer <id|#> <message>
/subagents spawn <agentId> <task> [--model <model>] [--thinking <level>]

/subagents info shows run metadata (status, timestamps, session id, transcript path, cleanup). Use sessions_history for a bounded, safety-filtered recall view; inspect the transcript path on disk when you need the raw full transcript.

Thread binding controls

These commands work on channels that support persistent thread bindings. See Thread supporting channels below.

/focus <subagent-label|session-key|session-id|session-label>
/unfocus
/agents
/session idle <duration|off>
/session max-age <duration|off>

Spawn behavior

/subagents spawn starts a background sub-agent as a user command (not an internal relay) and sends one final completion update back to the requester chat when the run finishes.

- The spawn command is non-blocking; it returns a run id immediately. - On completion, the sub-agent announces a summary/result message back to the requester chat channel. - Completion is push-based. Once spawned, do **not** poll `/subagents list`, `sessions_list`, or `sessions_history` in a loop just to wait for it to finish; inspect status only on-demand for debugging or intervention. - On completion, OpenClaw best-effort closes tracked browser tabs/processes opened by that sub-agent session before the announce cleanup flow continues. - OpenClaw tries direct `agent` delivery first with a stable idempotency key. - If direct delivery fails, it falls back to queue routing. - If queue routing is still not available, the announce is retried with a short exponential backoff before final give-up. - Completion delivery keeps the resolved requester route: thread-bound or conversation-bound completion routes win when available; if the completion origin only provides a channel, OpenClaw fills the missing target/account from the requester session's resolved route (`lastChannel` / `lastTo` / `lastAccountId`) so direct delivery still works. The completion handoff to the requester session is runtime-generated internal context (not user-authored text) and includes:

- `Result` — latest visible `assistant` reply text, otherwise sanitized latest tool/toolResult text. Terminal failed runs do not reuse captured reply text.
- `Status` — `completed successfully` / `failed` / `timed out` / `unknown`.
- Compact runtime/token stats.
- A delivery instruction telling the requester agent to rewrite in normal assistant voice (not forward raw internal metadata).

- `--model` and `--thinking` override defaults for that specific run. - Use `info`/`log` to inspect details and output after completion. - `/subagents spawn` is one-shot mode (`mode: "run"`). For persistent thread-bound sessions, use `sessions_spawn` with `thread: true` and `mode: "session"`. - For ACP harness sessions (Claude Code, Gemini CLI, OpenCode, or explicit Codex ACP/acpx), use `sessions_spawn` with `runtime: "acp"` when the tool advertises that runtime. See [ACP delivery model](/tools/acp-agents#delivery-model) when debugging completions or agent-to-agent loops. When the `codex` plugin is enabled, Codex chat/thread control should prefer `/codex ...` over ACP unless the user explicitly asks for ACP/acpx. - OpenClaw hides `runtime: "acp"` until ACP is enabled, the requester is not sandboxed, and a backend plugin such as `acpx` is loaded. `runtime: "acp"` expects an external ACP harness id, or an `agents.list[]` entry with `runtime.type="acp"`; use the default sub-agent runtime for normal OpenClaw config agents from `agents_list`.

Context modes

Native sub-agents start isolated unless the caller explicitly asks to fork the current transcript.

Mode	When to use it	Behavior
isolated	Fresh research, independent implementation, slow tool work, or anything that can be briefed in the task text	Creates a clean child transcript. This is the default and keeps token use lower.
fork	Work that depends on the current conversation, prior tool results, or nuanced instructions already present in the requester transcript	Branches the requester transcript into the child session before the child starts.

Use fork sparingly. It is for context-sensitive delegation, not a replacement for writing a clear task prompt.

Tool: sessions_spawn

Starts a sub-agent run with deliver: false on the global subagent lane, then runs an announce step and posts the announce reply to the requester chat channel.

Availability depends on the caller's effective tool policy. The coding and full profiles expose sessions_spawn by default. The messaging profile does not; add tools.alsoAllow: ["sessions_spawn", "sessions_yield", "subagents"] or use tools.profile: "coding" for agents that should delegate work. Channel/group, provider, sandbox, and per-agent allow/deny policies can still remove the tool after the profile stage. Use /tools from the same session to confirm the effective tool list.

Defaults:

• 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.
• 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.
• Run timeout: if sessions_spawn.runTimeoutSeconds is omitted, OpenClaw uses agents.defaults.subagents.runTimeoutSeconds when set; otherwise it falls back to 0 (no timeout).

Tool parameters

The task description for the sub-agent. Optional human-readable label. Spawn under another agent id when allowed by `subagents.allowAgents`. `acp` is only for external ACP harnesses (`claude`, `droid`, `gemini`, `opencode`, or explicitly requested Codex ACP/acpx) and for `agents.list[]` entries whose `runtime.type` is `acp`. ACP-only. Resumes an existing ACP harness session when `runtime: "acp"`; ignored for native sub-agent spawns. ACP-only. Streams ACP run output to the parent session when `runtime: "acp"`; omit for native sub-agent spawns. Override the sub-agent model. Invalid values are skipped and the sub-agent runs on the default model with a warning in the tool result. Override thinking level for the sub-agent run. Defaults to `agents.defaults.subagents.runTimeoutSeconds` when set, otherwise `0`. When set, the sub-agent run is aborted after N seconds. When `true`, requests channel thread binding for this sub-agent session. If `thread: true` and `mode` omitted, default becomes `session`. `mode: "session"` requires `thread: true`. `"delete"` archives immediately after announce (still keeps the transcript via rename). `require` rejects spawn unless the target child runtime is sandboxed. `fork` branches the requester's current transcript into the child session. Native sub-agents only. Use `fork` only when the child needs the current transcript. `sessions_spawn` does **not** accept channel-delivery params (`target`, `channel`, `to`, `threadId`, `replyTo`, `transport`). For delivery, use `message`/`sessions_send` from the spawned run.

Thread-bound sessions

When thread bindings are enabled for a channel, a sub-agent can stay bound to a thread so follow-up user messages in that thread keep routing to the same sub-agent session.

Thread supporting channels

Discord is currently the only supported channel. It supports persistent thread-bound subagent sessions (sessions_spawn with thread: true), manual thread controls (/focus, /unfocus, /agents, /session idle, /session max-age), and adapter keys channels.discord.threadBindings.enabled, channels.discord.threadBindings.idleHours, channels.discord.threadBindings.maxAgeHours, and channels.discord.threadBindings.spawnSubagentSessions.

Quick flow

`sessions_spawn` with `thread: true` (and optionally `mode: "session"`). OpenClaw creates or binds a thread to that session target in the active channel. Replies and follow-up messages in that thread route to the bound session. Use `/session idle` to inspect/update inactivity auto-unfocus and `/session max-age` to control the hard cap. Use `/unfocus` to detach manually.

Manual controls

Command	Effect
/focus <target>	Bind the current thread (or create one) to a sub-agent/session target
/unfocus	Remove the binding for the current bound thread
/agents	List active runs and binding state (thread:<id> or unbound)
/session idle	Inspect/update idle auto-unfocus (focused bound threads only)
/session max-age	Inspect/update hard cap (focused bound threads only)

Config switches

• Global default: session.threadBindings.enabled, session.threadBindings.idleHours, session.threadBindings.maxAgeHours.
• Channel override and spawn auto-bind keys are adapter-specific. See Thread supporting channels above.

See Configuration reference and Slash commands for current adapter details.

Allowlist

List of agent ids that can be targeted via explicit `agentId` (`["*"]` allows any). Default: only the requester agent. If you set a list and still want the requester to spawn itself with `agentId`, include the requester id in the list. Default target-agent allowlist used when the requester agent does not set its own `subagents.allowAgents`. Block `sessions_spawn` calls that omit `agentId` (forces explicit profile selection). Per-agent override: `agents.list[].subagents.requireAgentId`.

If the requester session is sandboxed, sessions_spawn rejects targets that would run unsandboxed.

Discovery

Use agents_list to see which agent ids are currently allowed for sessions_spawn. The response includes each listed agent's effective model and embedded runtime metadata so callers can distinguish PI, Codex app-server, and other configured native runtimes.

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.
• Auto-archive applies equally to depth-1 and depth-2 sessions.
• Browser cleanup is separate from archive cleanup: tracked browser tabs/processes are best-effort closed when the run finishes, even if the transcript/session record is kept.

Nested sub-agents

By default, sub-agents cannot spawn their own sub-agents (maxSpawnDepth: 1). Set maxSpawnDepth: 2 to enable one level of nesting — the orchestrator pattern: main → orchestrator sub-agent → worker sub-sub-agents.

{
  agents: {
    defaults: {
      subagents: {
        maxSpawnDepth: 2, // allow sub-agents to spawn children (default: 1)
        maxChildrenPerAgent: 5, // max active children per agent session (default: 5)
        maxConcurrent: 8, // global concurrency lane cap (default: 8)
        runTimeoutSeconds: 900, // default timeout for sessions_spawn when omitted (0 = no timeout)
      },
    },
  },
}

Depth levels

Depth	Session key shape	Role	Can spawn?
0	agent:<id>:main	Main agent	Always
1	agent:<id>:subagent:<uuid>	Sub-agent (orchestrator when depth 2 allowed)	Only if maxSpawnDepth >= 2
2	agent:<id>:subagent:<uuid>:subagent:<uuid>	Sub-sub-agent (leaf worker)	Never

Announce chain

Results flow back up the chain:

1. Depth-2 worker finishes → announces to its parent (depth-1 orchestrator).
2. Depth-1 orchestrator receives the announce, synthesizes results, finishes → announces to main.
3. Main agent receives the announce and delivers to the user.

Each level only sees announces from its direct children.

**Operational guidance:** start child work once and wait for completion events instead of building poll loops around `sessions_list`, `sessions_history`, `/subagents list`, or `exec` sleep commands. `sessions_list` and `/subagents list` keep child-session relationships focused on live work — live children remain attached, ended children stay visible for a short recent window, and stale store-only child links are ignored after their freshness window. This prevents old `spawnedBy` / `parentSessionKey` metadata from resurrecting ghost children after restart. If a child completion event arrives after you already sent the final answer, the correct follow-up is the exact silent token `NO_REPLY` / `no_reply`.

Tool policy by depth

• Role and control scope are written into session metadata at spawn time. That keeps flat or restored session keys from accidentally regaining orchestrator privileges.
• Depth 1 (orchestrator, when maxSpawnDepth >= 2): gets sessions_spawn, subagents, sessions_list, sessions_history so it can manage its children. Other session/system tools remain denied.
• Depth 1 (leaf, when maxSpawnDepth == 1): no session tools (current default behavior).
• Depth 2 (leaf worker): no session tools — sessions_spawn is always denied at depth 2. Cannot spawn further children.

Per-agent spawn limit

Each agent session (at any depth) can have at most maxChildrenPerAgent (default 5) active children at a time. This prevents runaway fan-out from a single orchestrator.

Cascade stop

Stopping a depth-1 orchestrator automatically stops all its depth-2 children:

• /stop in the main chat stops all depth-1 agents and cascades to their depth-2 children.
• /subagents kill <id> stops a specific sub-agent and cascades to its children.
• /subagents kill all stops all sub-agents for the requester and cascades.

Authentication

Sub-agent auth is resolved by agent id, not by session type:

• The sub-agent session key is agent:<agentId>:subagent:<uuid>.
• The auth store is loaded from that agent's agentDir.
• The main agent's auth profiles are merged in as a fallback; agent profiles override main profiles on conflicts.

The merge is additive, so main profiles are always available as fallbacks. Fully isolated auth per agent is not supported yet.

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.
• If the latest assistant text is the exact silent token NO_REPLY / no_reply, announce output is suppressed even if earlier visible progress existed.

Delivery depends on requester depth:

• Top-level requester sessions use a follow-up agent call with external delivery (deliver=true).
• Nested requester subagent sessions receive an internal follow-up injection (deliver=false) so the orchestrator can synthesize child results in-session.
• If a nested requester subagent session is gone, OpenClaw falls back to that session's requester when available.

For top-level requester sessions, completion-mode direct delivery first resolves any bound conversation/thread route and hook override, then fills missing channel-target fields from the requester session's stored route. That keeps completions on the right chat/topic even when the completion origin only identifies the channel.

Child completion aggregation is scoped to the current requester run when building nested completion findings, preventing stale prior-run child outputs from leaking into the current announce. Announce replies preserve thread/topic routing when available on channel adapters.

Announce context

Announce context is normalized to a stable internal event block:

Field	Source
Source	subagent or cron
Session ids	Child session key/id
Type	Announce type + task label
Status	Derived from runtime outcome (success, error, timeout, or unknown) — not inferred from model text
Result content	Latest visible assistant text, otherwise sanitized latest tool/toolResult text
Follow-up	Instruction describing when to reply vs stay silent

Terminal failed runs report failure status without replaying captured reply text. On timeout, if the child only got through tool calls, announce can collapse that history into a short partial-progress summary instead of replaying raw tool output.

Stats line

Announce payloads include a stats line at the end (even when wrapped):

• Runtime (e.g. runtime 5m12s).
• Token usage (input/output/total).
• Estimated cost when model pricing is configured (models.providers.*.models[].cost).
• sessionKey, sessionId, and transcript path so the main agent can fetch history via sessions_history or inspect the file on disk.

Internal metadata is meant for orchestration only; user-facing replies should be rewritten in normal assistant voice.

Why prefer sessions_history

sessions_history is the safer orchestration path:

• Assistant recall is normalized first: thinking tags stripped; <relevant-memories> / <relevant_memories> scaffolding stripped; plain-text tool-call XML payload blocks (<tool_call>, <function_call>, <tool_calls>, <function_calls>) stripped, including truncated payloads that never close cleanly; downgraded tool-call/result scaffolding and historical-context markers stripped; leaked model control tokens (<|assistant|>, other ASCII <|...|>, full-width <｜...｜>) stripped; malformed MiniMax tool-call XML stripped.
• Credential/token-like text is redacted.
• Long blocks can be truncated.
• Very large histories can drop older rows or replace an oversized row with [sessions_history omitted: message too large].
• Raw on-disk transcript inspection is the fallback when you need the full byte-for-byte transcript.

Tool policy

Sub-agents use the same profile and tool-policy pipeline as the parent or target agent first. After that, OpenClaw applies the sub-agent restriction layer.

With no restrictive tools.profile, sub-agents get all tools except session tools and system tools:

• sessions_list
• sessions_history
• sessions_send
• sessions_spawn

sessions_history remains a bounded, sanitized recall view here too — it is not a raw transcript dump.

When maxSpawnDepth >= 2, depth-1 orchestrator sub-agents additionally receive sessions_spawn, subagents, sessions_list, and sessions_history so they can manage their children.

Override via config

{
  agents: {
    defaults: {
      subagents: {
        maxConcurrent: 1,
      },
    },
  },
  tools: {
    subagents: {
      tools: {
        // deny wins
        deny: ["gateway", "cron"],
        // if allow is set, it becomes allow-only (deny still wins)
        // allow: ["read", "exec", "process"]
      },
    },
  },
}

tools.subagents.tools.allow is a final allow-only filter. It can narrow the already-resolved tool set, but it cannot add back a tool removed by tools.profile. For example, tools.profile: "coding" includes web_search/web_fetch but not the browser tool. To let coding-profile sub-agents use browser automation, add browser at the profile stage:

{
  tools: {
    profile: "coding",
    alsoAllow: ["browser"],
  },
}

Use per-agent agents.list[].tools.alsoAllow: ["browser"] when only one agent should get browser automation.

Concurrency

Sub-agents use a dedicated in-process queue lane:

• Lane name: subagent
• Concurrency: agents.defaults.subagents.maxConcurrent (default 8)

Liveness and recovery

OpenClaw does not treat endedAt absence as permanent proof that a sub-agent is still alive. Unended runs older than the stale-run window stop counting as active/pending in /subagents list, status summaries, descendant completion gating, and per-session concurrency checks.

After a gateway restart, stale unended restored runs are pruned unless their child session is marked abortedLastRun: true. Those restart-aborted child sessions remain recoverable through the sub-agent orphan recovery flow, which sends a synthetic resume message before clearing the aborted marker.

If a sub-agent spawn fails with Gateway `PAIRING_REQUIRED` / `scope-upgrade`, check the RPC caller before editing pairing state. Internal `sessions_spawn` coordination should connect as `client.id: "gateway-client"` with `client.mode: "backend"` over direct loopback shared-token/password auth; that path does not depend on the CLI's paired-device scope baseline. Remote callers, explicit `deviceIdentity`, explicit device-token paths, and browser/node clients still need normal device approval for scope upgrades.

Stopping

• Sending /stop in the requester chat aborts the requester session and stops any active sub-agent runs spawned from it, cascading to nested children.
• /subagents kill <id> stops a specific sub-agent and cascades to its children.

Limitations

• Sub-agent announce is best-effort. If the gateway restarts, pending "announce back" work is lost.
• Sub-agents still share the same gateway process resources; treat maxConcurrent as a safety valve.
• sessions_spawn is always non-blocking: it returns { status: "accepted", runId, childSessionKey } immediately.
• Sub-agent context only injects AGENTS.md + TOOLS.md (no SOUL.md, IDENTITY.md, USER.md, HEARTBEAT.md, or BOOTSTRAP.md).
• Maximum nesting depth is 5 (maxSpawnDepth range: 1–5). Depth 2 is recommended for most use cases.
• maxChildrenPerAgent caps active children per session (default 5, range 1–20).

Related

• ACP agents
• Agent send
• Background tasks
• Multi-agent sandbox tools