Anatomy of a run bundle

#The unit that makes the rest possible

A run bundle is the single artifact that captures one execution of your agent. Everything else — replay, audit, human gates, cost attribution — is downstream of getting this one object right. Get the bundle wrong and the rest is dashboards; get it right and production cognition becomes something you can hand to another engineer without folklore.

#A frozen workflow

The bundle opens with a DAG snapshot: the workflow exactly as it existed when the run started — steps, gate policies, edges — not as it looks now. Freezing it is the whole point. A replay three months later reconstructs the same structure even after you've reshaped the live workflow, so "what the agent thought was running" is never in question.

#Typed trace events, not logs

Every meaningful moment is recorded as a typed fact: step_started/step_completed, model_io (provider, model, tokens, cost, duration), tool_call, human_gate_resolved, artifact_written. These aren't free-text lines you regex later — they're structured events you can query, diff, and correlate to DAG structure.

#Artifacts with schemas

The outputs each step produced are validated against a registered schema and versioned. That's what lets you diff artifact v1 against v2 and prove exactly which revision crossed an approval boundary.

#Gates: humans on purpose

PAUSE_BEFORE, PAUSE_AFTER, CHECKPOINT. When a human approves, rejects, or edits, the decision is recorded inside the bundle, attached to the structured output it gated — not stranded in a Slack thread that no one can cryptographically tie to what the customer received.

#Cost, attributed

Because model_io carries tokens and cost per step, invoice lines stop being a mystery: spend ties to the specific call that justified it, against the specific artifact revision it produced.

#The integrity hash

Finally, a sha256 over the whole bundle. Forward it to a reviewer, a customer, or a regulator and they can verify it wasn't altered in transit. That single property is the difference between a dashboard scrape and evidence that survives forwarding.

#Why bundle it at all

Because the alternative is forensic improvisation — pasting logs into spreadsheets, correlating timestamps across vendors, debating whether two "similar" traces meant the same user-visible outcome. A bundle turns each of those into an ordinary engineering problem: diff the JSON, verify the hash, correlate the trace ids.