Skip to main content

Multi-Agent Synthesis Fix

Date: 2026-01-22 Status: 🟢 In-use Author: Sri Aradhyula <sraradhy@cisco.com>

Summary​

Fixed an issue where the supervisor's final answer only contained the last sub-agent's response instead of a synthesized summary from all agents when multiple agents were invoked.

Problem​

When a user query required multiple agents (e.g., "get apps for CAIPE, search for CAIPE in RAG, look for info for cnoe-io/ai-platform-engineering"), the system would:

  1. Stream content from ALL agents correctly (user sees all results)
  2. But return only the LAST agent's response as the final answer

Root Cause​

The sub_agent_complete boolean flag in StreamState was blocking all streaming after the first sub-agent completed:

# OLD CODE - PROBLEMATIC
if state.sub_agent_complete:
    logger.info("🛑 Skipping streaming chunk - sub-agent already sent complete_result")
    return  # Blocks ALL subsequent streaming!

This caused:

  1. First agent completes → sub_agent_complete = True
  2. Second agent's streaming chunks → SKIPPED
  3. Third agent's streaming chunks → SKIPPED
  4. Supervisor's synthesis → SKIPPED

The _get_final_content() method would then return only the first agent's content (from artifact-updates), not the supervisor's synthesis.

Solution​

1. Changed from Boolean to Counter​

# OLD
sub_agent_complete: bool = False

# NEW
sub_agents_completed: int = 0  # Track count for multi-agent scenarios

2. Removed Streaming Block​

Instead of blocking ALL streaming after first agent completes, we now continue accumulating content:

# OLD
if state.sub_agent_complete:
    logger.info("🛑 Skipping streaming chunk - sub-agent already sent complete_result")
    return

# NEW
# NOTE: We no longer block streaming after sub-agent completion.
# For multi-agent scenarios, the supervisor needs to synthesize results
# from all sub-agents, so we must continue accumulating content.

3. Updated Final Content Priority​

_get_final_content() now prioritizes differently based on scenario:

def _get_final_content(self, state: StreamState) -> tuple:
    """
    Priority order for multi-agent scenarios:
    1. Sub-agent DataPart (structured data - e.g., Jarvis forms)
    2. Supervisor content (synthesis from multiple agents)
    3. Sub-agent text content (single agent fallback)
    """
    if state.sub_agent_datapart:
        return state.sub_agent_datapart, True

    # Multi-agent scenario: prefer supervisor synthesis
    if state.sub_agents_completed > 1 and state.supervisor_content:
        raw_content = ''.join(state.supervisor_content)
        return self._extract_final_answer(raw_content), False

    # Single agent: use sub-agent content
    if state.sub_agent_content:
        raw_content = ''.join(state.sub_agent_content)
        return self._extract_final_answer(raw_content), False

    # Fallback to supervisor content
    if state.supervisor_content:
        raw_content = ''.join(state.supervisor_content)
        return self._extract_final_answer(raw_content), False

    return '', False

Behavior Changes​

ScenarioBeforeAfter
Single agentSub-agent's response used ✓Sub-agent's response used ✓
Multiple agentsLast agent's response only ✗Supervisor synthesis used ✓
DataPart (e.g., Jarvis forms)DataPart used ✓DataPart used ✓

Files Changed​

  • ai_platform_engineering/multi_agents/platform_engineer/protocol_bindings/a2a/agent_executor.py

Testing​

Tested with query: "get apps for CAIPE, search for CAIPE in RAG, look for info for cnoe-io/ai-platform-engineering"

Expected result: Final answer contains synthesized summary from all three agents (ArgoCD, RAG, GitHub).