§ 05 · Consumption
The consumption layer is what the agent produces: reports, alerts, dashboards, API calls, events, notebook artifacts. It is also where agentic systems become operational rather than advisory, the moment they stop returning answers and start triggering actions.
§ Role in the stack
Traditional analytics consumption means dashboards and reports. Agentic analytics expands this significantly. The same agent that answers "why did revenue come in short?" can, given the right action surface, open a ticket for the account team, draft a mitigation plan for review, or push a notification into the channel where the VP lives. Consumption is no longer just rendering; it is dispatch.
Two design decisions shape this layer. First, the artifact format: agents should produce structured, replayable results, not text blobs. A notebook, a JSON payload with a schema, a rendered report with attached queries: anything that lets a human inspect the reasoning and a tool replay the result. Second, the action surface: what the agent is allowed to do on behalf of the user, expressed through bounded, audited tool calls rather than raw credentials.
Getting this layer right is what moves an agent from advisory (stage 5) to operationally autonomous (stage 6). Getting it wrong creates surface area for agent mistakes to become production incidents.
§ Open standards
Consumption has fewer agentic-specific standards than the layers below it, because the substrates (APIs, events, notebooks) predate agents. The relevant question is which of those substrates your agents should emit into.
The declarative format for the APIs the agent calls to deliver results: notification services, ticketing systems, CRM updates, workflow triggers. If actions live behind OpenAPI specs, most agent frameworks can register them as tools without custom glue.
Specification →The event-driven counterpart to OpenAPI. Matters for consumption because agent output increasingly lands on streaming topics (Kafka, NATS, SNS) rather than synchronous APIs. AsyncAPI specs document what the agent emits and what downstream consumers expect.
Specification →A standard envelope for event payloads the agent emits. Used together with AsyncAPI for the payload schema. Keeps agent-emitted events interoperable across event buses, serverless platforms, and notification services.
Specification →The notebook file format. Underrated as a consumption standard: if the agent produces a notebook that contains its plan, queries, intermediate results, and narrative, the artifact is immediately inspectable by humans and replayable by other tools.
Specification →§ Failure modes
Opaque result formats. The agent emits a paragraph of natural language. Downstream systems cannot parse it, humans cannot easily compare it to prior answers, and there is no structured way to say "this conclusion was based on these queries." Fix: every agent output carries structured metadata alongside the narrative (queries run, tables touched, confidence signals, versioned semantic definitions referenced).
Unreplayable audit trails. The notification goes out. Two days later someone asks, "what did the agent actually conclude, and what evidence did it use?" If the consumption artifact is a chat message, the answer is gone. Fix: persist the full artifact (notebook, structured report, traced run) at emission time, not reconstruct it from logs.
No action semantics. The agent says "I recommend pausing the campaign." Nothing pauses. The user has to translate the recommendation into an action in another system. Either stop framing output as recommendations, or give the agent a bounded action surface (OpenAPI-defined tools, dry-run modes, policy gates) so recommendations can be enacted with a review step.
Consumption that does not close the loop. The agent fires a notification. The user acts on it. The outcome never returns to the agent. Over time, this makes evaluation impossible: you cannot tell which recommendations worked. Fix: emit outcomes back into the event stream (CloudEvents over AsyncAPI) and treat them as training and evaluation signal.