Agent output provenance — tracking trust and sources across multi-agent conversations

Context

AutoGen orchestrates multi-agent conversations where agents generate content collaboratively. When Agent A generates a claim, Agent B refines it, and Agent C produces the final output — the provenance chain is lost. The final output carries no metadata about:

• Which agent contributed which part
• What sources each agent used
• The confidence level of each claim
• Whether any part was human-reviewed

Why this matters

1. Multi-agent trust — in a GroupChat, not all agents are equally reliable. An agent sourcing from official APIs (high trust) is different from one doing unverified inference (low trust). The final output should reflect this.

2. Debugging — when a multi-agent output contains errors, tracing back to which agent introduced the error requires provenance.

3. EU AI Act Article 50 (August 2, 2026) — AI-generated content must carry transparency metadata. Multi-agent outputs are AI-generated content.

Possible approach

Each message in the conversation could carry lightweight provenance:

message = {
    "content": "Revenue was $4.2B",
    "role": "assistant",
    "name": "analyst_agent",
    # provenance fields:
    "trust_score": 0.85,
    "source_tier": 2,  # verified secondary source
    "sources": ["SEC 10-Q filing"],
    "ai_generated": True,
    "verified": False,
}

The final output aggregates trust scores from contributing messages — lower of contributing trust levels, not average.

Prior art

• AKF defines a compact provenance schema (~15 tokens of JSON) for this
• C2PA handles media provenance but is too heavy for message-level metadata
• LangChain has callback handlers but no trust scoring

Has the AutoGen team considered output provenance? Especially as multi-agent systems move toward production use cases where auditability matters.