[feat] Observable context/state for inter-agent communication. #338

Activity

1. 0xbrainkid commented on Apr 15, 2026

   

   0xbrainkid

   on Apr 15, 2026

   More actions

   Observable context and state for inter-agent communication is the right foundation for making multi-agent pipelines debuggable — if you cannot see what context an agent is operating on when it makes a decision, you cannot reason about why it made that decision.

   One extension worth building into the observable state design from the start: identity provenance on context elements. In a multi-agent AgentStack pipeline, context accumulates from multiple sources — the user, previous agents, tool calls, external APIs. Without provenance tracking, a downstream agent cannot distinguish context that came from a trusted internal agent vs. context injected by an external tool or user message.

   A minimal addition to observable context:

   @dataclass
   class ContextElement:
       key: str
       value: Any
       source_agent_id: str | None = None       # which agent produced this
       source_trust_score: float | None = None  # trust at time of production
       source_type: Literal["agent", "tool", "user", "external"] = "agent"
       timestamp: datetime = field(default_factory=datetime.utcnow)

   With this, observable state becomes observable AND attributable — a downstream agent can see "this context element came from a high-trust internal agent" vs. "this came from an external API with no trust record." This is the difference between observable state for debugging and observable state for safe multi-agent operation.

   AgentFolio provides source_trust_score for any verified agent identity — plugging the lookup into context observation would make AgentStack pipelines trust-aware without requiring changes to the core agent logic.