Turn session telemetry into a readable story, action items, and guardrail recommendations.
You already have raw runtime events, guardrail decisions, approvals, and receipts. Runtime intelligence adds the product layer on top: it turns a session into a short summary, a grouped story timeline, follow-up action items, and guardrail recommendations you can review in the dashboard.When it is ready, you can answer “what happened, why did it matter, and what should we change next?” without reading every event by hand.
Runtime intelligence is derived from telemetry. Raw events, guardrail decisions, approvals, receipts, declared capabilities, and boundary records remain the source of truth.
A plain-language title, outcome, primary actor, involved tools/resources/environments, risk level, and confidence
Event ids from the session
Story timeline
A compressed timeline of meaningful steps, not one card per event unless the event matters on its own
Event ids, plus related guardrail decision, approval, and receipt ids when present
Action items
Concrete next steps such as reviewing an unknown tool, declaring a capability, requiring approval, or fixing metadata
Evidence event ids and a confidence label
Guardrail recommendations
Suggested monitor, warning, approval, or block policies for the observed behavior
Evidence event ids, affected tool/action/resource/environment, and any matching template key
Runtime intelligence can return no action items. For example, a harmless read-only session should produce a readable story without telling you to do unnecessary work.
Runtime intelligence is generated for sessions with events. It is usually queued when a session completes, when a run completes, when an error or failed action occurs, when an approval is requested or resolved, or when Apie observes a high-signal event such as:
A guardrail decision that warns, blocks, or requires approval
An unknown action or resource
A production create, update, delete, or execute action
Access to risky resources such as secrets, pipeline runs, deployment events, or database records
While generation is running, the dashboard can show Story generation queued or Generating session story. When the result validates, the story replaces the deterministic fallback view.If new events arrive after a result was generated, the dashboard marks the story as Newer events are available and keeps the previous result visible until a newer result is ready. Workspace managers can regenerate a stale or failed story from the session replay page.
The raw event stream stays available for audit and debugging. Runtime intelligence reads a compact, redacted context built from those events and produces a validated view for the dashboard.Each useful claim must point back to evidence:
Field
Why it matters
evidenceEventIds
Lets you jump from a summary, action item, or recommendation back to the events that support it
sourceBasis
Distinguishes declared, observed, inferred, unknown, and violated behavior
confidence
Shows whether the interpretation is high confidence or should be treated cautiously
contextWarnings
Surfaces limits in the available context, such as compacted or incomplete evidence
The generated story may infer meaning from event labels, metadata, and guardrail context, but it should use “unknown” when the evidence is not strong enough. It does not overwrite event metadata or boundary records.
Guardrails decide what happens at runtime. Runtime intelligence explains what happened after the fact and may recommend what to guard next.
Guardrail concept
Runtime intelligence behavior
Existing guardrail decision
Timeline cards can reference the decision that allowed, warned, blocked, or required approval
Approval request
Timeline cards can link the approval that paused or resolved the action
Recommended guardrail
The dashboard can suggest a mode such as Monitor, Warn, Require approval, or Block
Declared capabilities
Action items can suggest declaring expected tools or fixing missing action/resource metadata
A recommendation is not an active policy. Review it against your production intent, then enable the matching template or create the guardrail you want.
Runtime intelligence is designed for review, not certification. It should not claim that an agent is safe, compliant, certified, fully governed, or fully controlled.Before model generation, Apie builds a compact context from canonical fields, redacted metadata summaries, evidence ids, guardrail references, approvals, receipts, declared tools, boundary state, and agent profile data. If sensitive content is still detected in the final model input, generation fails instead of sending that input.
