Designing Inspectable Artifacts

An agent that runs without leaving an artifact is, for review purposes, a rumour. The exam expects you to know which artifacts to produce, which properties make them actually inspectable, and which traps make them theatre.

The minimum artifact bundle

For any side-effecting run, the agent should leave behind:

1. The plan — the structured plan that was approved at the boundary.
2. The trace — every model call and tool call, joined by a stable trace ID.
3. The diff (or equivalent output) — what actually changed in the world.
4. The outcome — success / failure with a classified reason.

The canonical packaging for a code-modifying agent is a pull request: plan in the description, diff in the files, tool calls and trace linked from the body.

Properties that make an artifact inspectable

| Property | Why it matters | | --- | --- | | Durable | Outlives the agent's session, the deploy, the on-call rotation. | | Addressable | Has a stable URL or ID a reviewer can paste into Slack two weeks later. | | Granular | Line-by-line, step-by-step — coarse summaries hide the bugs. | | Reproducible | The inputs are captured well enough to re-run the agent. | | Accessible | Reviewers have permission to read it without a ticket. |

Failure-path artifacts matter more

A common trap: the agent emits gorgeous artifacts on the happy path and error: 500 on failure. This is exactly backwards. On failure you need more detail, not less:

• the last successful step,
• the tool call that broke,
• the classified error (permission_denied / validation_failed / transient),
• the partial plan up to the failure,
• a one-line "what would I need to retry?" hint.

Exam tip: when a question contrasts two designs, the one that produces a richer failure artifact is almost always the right answer.

The trace ID is the join key

Every artifact carries the same trace ID. That ID is what joins:

plan_v1 ── trace_id ──> tool_call_1 ── trace_id ──> diff ── trace_id ──> deploy_log

Without a stable trace ID you have files, not an investigation.

Quick check

Where this shows up on the exam

Several Domain 1 questions phrase inspectability as "the next morning". Picture a reviewer arriving with a coffee, opening Slack, and trying to understand what the agent did overnight. If your design lets them do that in under five minutes — plan, diff, trace, classified outcome — you are answering the question correctly.