Skip to main content

Artifacts and Replay

To make agentic systems reliable, you need evidence:

  • what happened,
  • why it happened,
  • and what inputs led to that outcome.

ARP and JARVIS treat these as first-class artifacts.

Events (NDJSON)​

Execution produces RunEvents (and sometimes NodeRun-scoped events). In the JARVIS stack:

  • events are appended durably to an Event Stream store,
  • and served back as NDJSON streams for efficient tailing and tooling.

NDJSON is used because it works well for:

  • streaming over plain HTTP,
  • incremental parsing,
  • log-like storage and replay.

Artifacts (blobs + references)​

Some outputs are too large or too structured to inline into Run/NodeRun payloads.

JARVIS uses an Artifact Store to:

  • store blobs (bytes),
  • return stable references (ArtifactRef) that can be recorded in NodeRun outputs or RunEvents.

Replay​

Replay is the ability to re-run (or “simulate”) execution in a controlled way:

  • deterministically re-bind to previously selected candidates,
  • compare outputs across runs,
  • validate that changes to planners/selectors do not regress capability.

Even before full “replay mode” exists, durable events + artifacts enable:

  • debugging,
  • incident review,
  • evaluation pipelines.

Where to go next​