Skip to main content

Oversight Operating Model

This guide explains how to apply:

  • /docs/standards/human-oversight-standards in daily engineering operations without creating “approval theater”.

Objectives

An effective oversight model:

  • routes human attention to the highest-risk agent actions
  • reduces rubber-stamp approvals
  • preserves speed for low-risk work
  • creates clear records for later audit and learning

Inputs you need

Before this works well, have:

  • Agent Inventory (who/what/where)
  • Checkpoint Register (what actions require what oversight)
  • a simple risk classification method (see HOR-002)

Oversight levels in practice

Use the oversight levels described in /docs/standards/human-oversight-standards as an organization-wide vocabulary.

Operational guidance:

  • prefer Notify for low-risk, reversible changes
  • use Approval required for changes with high blast radius or boundary impact
  • use Supervised for time-sensitive or complex production workflows
  • use Human-executed for exceptionally sensitive actions (e.g., access boundary changes in high-risk environments)

Building the Checkpoint Register

Start small: define checkpoints for a short list of high-risk action categories.

Suggested initial categories:

  • schema migrations
  • auth/authz changes
  • tenancy logic and RLS/policy changes
  • secrets/credentials scope changes
  • CI/CD pipeline modifications
  • production deploys and rollbacks
  • dependency upgrades that affect runtime or security posture
  • environment configuration changes

Checkpoint record fields (recommended):

  • checkpoint name
  • action category
  • required oversight level
  • escalation criteria
  • evidence required (logs, diff, tests, rationale)
  • owner (team)

Risk classification (simple but effective)

A lightweight classification can be computed with 5 questions:

  1. Blast radius: how many systems/users/tenants can be affected?
  2. Reversibility: can you roll back safely and quickly?
  3. Boundary impact: does it touch auth, tenancy, secrets, or environment separation?
  4. Schema impact: does it change schema/migrations?
  5. Production impact: does it change production behavior directly?

Operational rule:

  • if 2+ answers are “high”, treat as high-risk and elevate oversight.

TODO: If you want formal scoring bands, define them in Framework and keep the standards “policy-agnostic”.

Approval packets (reduce rubber-stamping)

When an approval is required, the approver should receive a standard packet:

  • concise summary (what changes, why)
  • diff or change set reference
  • risk classification result
  • rationale for significant decisions
  • validation evidence (tests, checks, migration rehearsal, etc.)
  • rollback plan (for production-impacting changes)

If approvers lack this context, approvals become symbolic.

Escalation and emergency stops

Operationalize:

  • who can override an agent decision
  • who can trigger emergency stop
  • what happens immediately after (triage, containment, replay reconstruction)
  • how post-incident learning updates checkpoints and policies

Reference:

  • Evidence requirements live in /docs/standards/human-oversight-standards
  • Replayability support: /docs/framework/behavior-verification

Metrics that reveal whether oversight is working

Track a small set:

  • approval latency (median / p95)
  • override frequency and reasons
  • emergency stop events (count and outcomes)
  • incident rate tied to agentic changes
  • % of significant actions with complete rationale records

Interpretation tips:

  • very low latency + low context packets often indicates rubber-stamping
  • very high latency suggests checkpoints are too broad or poorly routed

Common failure modes (and mitigations)

  • Approval theater: approvals exist but no one understands them
    • mitigate with approval packets + sampling review
  • Checkpoint sprawl: too many approvals, everything slows
    • mitigate by narrowing to boundary-impacting actions
  • Shadow automation: teams bypass governed agents because governance is painful
    • mitigate by making compliance easier than bypassing

TODO: Add a sample checkpoint register table once you standardize your internal fields.