agent_deepagent: dispatch multi-task tool calls in parallel

[joshuadarron]

Problem

Deep Agent's task tool serializes multi-tool-call turns. When a Project Manager LLM emits N task(...) calls in a single assistant turn (idiomatic LangChain/LangGraph fan-out), the engine processes them one-by-one rather than concurrently.

Tracer evidence from a 28-minute coding-agent run: three engineer subagents marked for parallel execution by the PM ran strictly sequentially. op:enter/op:leave windows for agent_deepagent_subagent_eng1/2/3 did not overlap; wall time was eng1 + eng2 + eng3 instead of max(eng1, eng2, eng3). Estimated savings on that run: ~350 s (~20% of total).

Root cause

# nodes/agent_deepagent/deepagent.py:357-360
state = agent.invoke(
    {'messages': [HumanMessage(content=_safe_str(question.getPrompt() or ''))]},
    config={'callbacks': [_SSECallbackHandler(_send_sse)]},
)

agent.invoke() uses LangGraph's default tool-execution node, which iterates tool calls sequentially. The PM system prompt can tell the LLM to "issue all in a single message" but the engine still serializes them.

Proposed fix

Replace the default tool-execution node with one that fans out via asyncio.gather across tool_calls, or use LangGraph's Send API to spawn parallel branches per tool call. The existing _SSECallbackHandler is already async-safe; the SSE stream just needs per-task tagging so the UI doesn't interleave incoherently.

Acceptance

• Multi-task turns from a Deep Agent show overlapping op:enter windows in apaevt_flow traces.
• Single-task and other tools (shell/fs/git) are unaffected (they remain sequential by design).
• A regression test covering "PM emits 3 task calls" asserts wall time ≈ max-branch instead of sum-of-branches (with a tolerance).

Suggested labels

enhancement, performance, nodes/agent_deepagent

[github-actions]