Subagent Streaming

Prerequisite: Streaming basics — stream() API, EventType, StreamOptions, SSE integration. This page focuses on HarnessAgent’s child-agent event forwarding mechanism.

When a parent HarnessAgent.stream() triggers agent_spawn or agent_send, all intermediate events from the child agent are injected into the parent Flux<Event> in real time, tagged with an EventSource that identifies their origin — no extra configuration needed.

How It Works

AgentBase.createEventStream()
  │
  ├─ Creates FluxSink<Event>
  ├─ Constructs SubagentEventBus (impl = sink::next)
  └─ Injects bus into Reactor Context
        │
        ▼
  ReActAgent reasoning loop (runs inside Reactor Context)
        │
        ▼  acting phase calls agent_spawn
  AgentSpawnTool.execLocalSync()
        │
        ├─ Reads SubagentEventBus from Reactor Context
        ├─ Calls DefaultAgentManager.invokeAgentStream()
        │       │  child HarnessAgent.stream()
        │       │    each child Event ──map(withSource)──▶ tagged with EventSource
        │       └─ Flux<Event> flows back to execLocalSync
        │
        └─ doOnNext(bus::emit) ──▶ each child Event pushed into parent FluxSink live

Key constraint: the bus only exists in stream() mode. When using call(), no bus is present and agent_spawn falls back to the blocking invokeAgent path with no behaviour change.

Event Timeline

Caller
  └─ parent.stream()
        │
        ├─ REASONING(parent, chunk×N)   ← parent reasoning turn 1 (with tool_use)
        │
        │  [agent_spawn "researcher" starts]
        ├─ REASONING(child, chunk×M)    ← child reasoning (forwarded live, has EventSource)
        ├─ TOOL_RESULT(child, ...)      ← child tool result (forwarded live, if any)
        ├─ REASONING(child, last)       ← child final reasoning (forwarded live)
        ├─ AGENT_RESULT(child, last)    ← child final reply (forwarded live)
        │  [agent_spawn ends; child result returned as TOOL_RESULT to parent]
        │
        ├─ TOOL_RESULT(parent, ...)     ← parent receives child result
        ├─ REASONING(parent, chunk×K)   ← parent reasoning turn 2
        └─ AGENT_RESULT(parent, last)   ← parent final reply

Parent events have source == null; child events have source != null.

Identifying the Source with EventSource

Flux<Event> events = parent.stream(msgs, StreamOptions.defaults(), ctx);

events.subscribe(event -> {
    EventSource src = event.getSource();
    if (src == null) {
        // parent agent's own event
        System.out.printf("[parent][%s] %s%n",
                event.getType(), event.getMessage().getTextContent());
    } else {
        // child (or grandchild) agent event
        System.out.printf("[%s|depth=%d|path=%s][%s] %s%n",
                src.getAgentId(), src.getDepth(), src.getPath(),
                event.getType(), event.getMessage().getTextContent());
    }
});

EventSource Field Reference

Field	Meaning	Example
`agentKey`	Runtime instance handle; pass to `agent_send`	`agent:researcher:550e8400-…`
`agentId`	Child agent type ID (matches `subagents/<id>.md` filename)	`researcher`
`agentName`	Child agent display name (may be null)	`ResearcherAgent`
`sessionId`	Unique session ID for this child invocation	`sub-a1b2c3d4-…`
`parentSessionId`	Session ID of the parent agent	`sess-main-001`
`depth`	Nesting depth: direct child = 1, grandchild = 2, …	`1`
`path`	Slash-separated call hierarchy	`sess-main-001/researcher`
`taskId`	Reserved; `null` today (reserved for async streaming)	`null`

Path convention: <parentSessionId>/<agentId>. For deeper nesting it extends naturally: sess-001/planner/executor.

Multi-level Nesting (Grandchild Agents)

A child HarnessAgent injects its own bus at its stream() entry point. Grandchild events are first captured by the child bus (depth+1), then flow out through the child Flux<Event>, where AgentSpawnTool in the grandparent forwards them again to the grandparent bus. The path field accumulates depth automatically.

parent.stream()
  └─ child.stream()             depth=1, path="sess/planner"
        └─ grandchild.stream()  depth=2, path="sess/planner/executor"

Filter by depth or path prefix to target any level:

// Only depth-1 child REASONING events
events.filter(e -> e.getSource() != null
               && e.getSource().getDepth() == 1
               && e.getType() == EventType.REASONING)
      .subscribe(...);

// All events from any agent in the "executor" path segment
events.filter(e -> e.getSource() != null
               && e.getSource().getPath().contains("executor"))
      .subscribe(...);

Common Consumption Patterns

1. Real-time UI Rendering per Agent

Route events to separate UI panels based on source:

events.groupBy(e -> e.getSource() == null ? "parent" : e.getSource().getAgentId())
      .flatMap(group -> group.doOnNext(e -> renderToPanel(group.key(), e)))
      .subscribe();

2. Await Child Agent’s Final Reply

If you only care about the child’s result text (not intermediate events):

String childReply = events
        .filter(e -> e.getSource() != null
                  && "researcher".equals(e.getSource().getAgentId())
                  && e.getType() == EventType.AGENT_RESULT
                  && e.isLast())
        .map(e -> e.getMessage().getTextContent())
        .blockFirst();

3. Collect and Group by Agent

Map<String, List<Event>> byAgent = events.collectList().block()
        .stream()
        .collect(Collectors.groupingBy(e ->
                e.getSource() == null ? "parent" : e.getSource().getAgentId()));

4. SSE with Source Metadata

Forward child-agent origin to the browser:

@GetMapping(value = "/chat", produces = MediaType.TEXT_EVENT_STREAM_VALUE)
public Flux<ServerSentEvent<String>> chat(@RequestParam String message,
                                          @RequestParam String sessionId) {
    RuntimeContext ctx = RuntimeContext.builder().sessionId(sessionId).build();
    return agent.stream(
                    List.of(Msg.builder().role(MsgRole.USER).textContent(message).build()),
                    StreamOptions.defaults(), ctx)
            .map(event -> {
                Map<String, Object> payload = new LinkedHashMap<>();
                payload.put("type", event.getType());
                payload.put("text", event.getMessage().getTextContent());
                payload.put("last", event.isLast());
                if (event.getSource() != null) {
                    payload.put("agentId", event.getSource().getAgentId());
                    payload.put("depth",   event.getSource().getDepth());
                    payload.put("path",    event.getSource().getPath());
                }
                return ServerSentEvent.<String>builder()
                        .data(objectMapper.writeValueAsString(payload))
                        .build();
            });
}

Error Handling

When a child agent throws internally, the framework catches the exception and writes it as a TOOL_RESULT error string. The onError signal is not propagated to the parent Flux, so the parent stream continues normally. For fatal errors in the parent stream itself, standard Reactor semantics apply:

events.onErrorResume(e -> {
    log.error("Parent stream error", e);
    return Flux.empty();
}).subscribe(...);

Scope Boundaries

Scenario	Behaviour
`stream()` + sync local subagent (`timeout_seconds > 0`)	✔ Child events forwarded live with `EventSource`
`call()` mode (non-streaming)	✔ Normal blocking invocation; child result returned as `tool_result` string; no event forwarding
`timeout_seconds=0` async background task	✗ Streaming not yet supported (future extension point)
Remote subagent (Agent Protocol)	✗ Streaming not yet supported (future extension point)
Multi-level nesting (grandchild)	✔ Path/depth accumulate automatically; `FluxSink.next` is thread-safe

Streaming Basics — stream() API, EventType, full StreamOptions reference
Subagent — Subagent declarations, agent_spawn / agent_send parameters
Architecture — SubagentEventBus, Reactor Context injection, and StreamingHook lifecycle

Quick Start

Features

Multi Agent

Harness

How It Works

Event Timeline

Identifying the Source with EventSource

EventSource Field Reference

Multi-level Nesting (Grandchild Agents)

Common Consumption Patterns

1. Real-time UI Rendering per Agent

2. Await Child Agent’s Final Reply

3. Collect and Group by Agent

4. SSE with Source Metadata

Error Handling

Scope Boundaries

​How It Works

​Event Timeline

​Identifying the Source with EventSource

​EventSource Field Reference

​Multi-level Nesting (Grandchild Agents)

​Common Consumption Patterns

​1. Real-time UI Rendering per Agent

​2. Await Child Agent’s Final Reply

​3. Collect and Group by Agent

​4. SSE with Source Metadata

​Error Handling

​Scope Boundaries

​Related Documents

How It Works

Event Timeline

Identifying the Source with EventSource

EventSource Field Reference

Multi-level Nesting (Grandchild Agents)

Common Consumption Patterns

1. Real-time UI Rendering per Agent

2. Await Child Agent’s Final Reply

3. Collect and Group by Agent

4. SSE with Source Metadata

Error Handling

Scope Boundaries

Related Documents