A deployment runbook is easier for an agent to execute safely when the public surface is named in advance. The surface is the small set of routes, feeds, indexes, machine-readable files, and intentionally missing paths that define whether a release behaved correctly.

Without that inventory, an agent may verify only the page it edited. That catches syntax errors, but it can miss broken discovery pages, stale indexes, missing sitemap entries, or accidental exposure of routes that should not exist.

Write the inventory as a short checklist

The inventory does not need to describe the whole site. It should list the few public checks that matter for the current workflow:

  • the changed route or artifact
  • one listing, index, feed, or sitemap that should reference it
  • one machine-readable file if agents or crawlers depend on it
  • one negative route that should remain unavailable
  • the expected status code or visible marker for each check

This turns deployment verification into a concrete contract. The agent is not asked to "look around" or decide what seems important after the fact. It is given a small map of the public behavior that must be preserved.

Include absence as part of the surface

Public-surface inventories should include negative checks. For example, if drafts, diagnostics, private previews, or deprecated routes should not be reachable, list one representative path and the expected not-found response.

Absence checks are useful because they prevent a common automation blind spot: treating every successful response as good news. In publishing systems, a page that loads can still be a failure if it exposes content that was not meant to ship.

Keep the report small

A good final receipt can be brief:

Public surface inventory passed.
Changed route: 200 and expected heading found.
Discovery route: sitemap references the new item.
Machine-readable route: llms.txt regenerated and reachable.
Negative route: returned 404 as expected.

That receipt gives reviewers confidence without sharing private logs, local paths, account details, or environment-specific output. It also gives the next agent a reusable checklist for the same class of deployment.

The habit is simple: before promoting a change, name the public surface, verify each item, and report the result. Small inventories make agent deployments less dependent on improvisation and more resistant to silent regressions.