Skip to main content

Platform Engineer Streaming Architecture

Status: 🔴 Abandoned (Superseded by 2024-10-23-platform-engineer-streaming-architecture.md) Category: Architecture & Core Design Date: October 22, 2024

Current Status: ⚠️ Streaming Not Fully Working (Historical Note)​

Token-by-token streaming from sub-agents (like agent-komodor-p2p) to clients is currently NOT working due to LangGraph's tool execution model. This document explains why and outlines the solution path.

The Problem​

Current Architecture​

Client Request
    ↓
Platform Engineer (Deep Agent + LangGraph)
    ↓
A2ARemoteAgentConnectTool (blocks here!)
    ↓ (internally streams from sub-agent)
Sub-Agent streams response → Tool accumulates → Returns complete text
    ↓
Platform Engineer receives complete response as one chunk
    ↓
Client receives full response at once (no streaming)

Root Cause​

LangGraph tools are blocking by design. When Deep Agent invokes a tool:

  1. Tool execution blocks the graph
  2. A2ARemoteAgentConnectTool._arun() is called
  3. Inside _arun(), the tool DOES stream from the sub-agent via A2A protocol
  4. BUT it accumulates all chunks into accumulated_text
  5. Only returns the complete response when streaming finishes
  6. LangGraph receives this as a single ToolMessage

Code Evidence (ai_platform_engineering/utils/a2a_common/a2a_remote_agent_connect.py:198-226):

accumulated_text: list[str] = []

async for chunk in self._client.send_message_streaming(streaming_request):
    # Chunks ARE received from sub-agent
    writer({"type": "a2a_event", "data": chunk_dump})  # ← This writes somewhere but doesn't propagate

    if isinstance(chunk, A2ATaskArtifactUpdateEvent):
        text = extract_text(chunk)
        accumulated_text.append(text)  # ← Accumulating, not yielding!

# Return complete response after ALL chunks received
final_response = " ".join(accumulated_text).strip()
return Output(response=final_response)  # ← Blocking return

Testing Streaming​

Current State (Not Streaming)​

uvx --no-cache git+https://github.com/cnoe-io/agent-chat-cli.git a2a \
  --host 10.99.255.178 --port 8000

# Type: show me komodor clusters
#
# Behavior: Shows "Calling komodor..." → wait → complete response appears

After Fix (Streaming)​

# Same command
#
# Expected: Tokens appear one by one as they're generated by komodor agent