Skip to main content

Human Oversight Standards

These standards define requirements for human oversight of autonomous development agents.

Principle​

Humans remain ultimately accountable for agent actions. Oversight mechanisms MUST ensure humans can:

  • understand what agents are doing
  • intervene when necessary
  • maintain meaningful control
  • learn from agent behavior

Oversight levels (normative model)​

Organizations MUST define and use an oversight model. The following levels are RECOMMENDED:

Level 0: Autonomous​

Agent operates independently within defined boundaries.

Level 1: Notify​

Agent acts autonomously but notifies humans of actions.

Level 2: Approval required​

Agent proposes actions and waits for human approval.

Level 3: Supervised​

Human actively monitors the agent during execution.

Level 4: Human-executed​

Agent provides recommendations; a human executes.

Note: Examples and operating guidance for applying these levels belong in /docs/framework/.

Oversight requirements (HOR)​

HOR-001: Checkpoint definition​

Organizations MUST define oversight checkpoints for agent workflows.

Checkpoint definitions MUST include:

  • checkpoint name and purpose
  • which workflows/actions the checkpoint applies to
  • required oversight level at the checkpoint
  • escalation criteria

Checkpoint definitions MUST be reviewed periodically (at an organization-defined cadence).

HOR-002: Risk classification mapping​

Organizations MUST define a risk classification method and map classifications to oversight levels.

The classification method MUST consider:

  • reversibility
  • blast radius
  • access boundary impact (auth/tenancy/data access)
  • schema/migration impact
  • production impact

HOR-003: Approval capture​

When approvals are required, approval records MUST include:

  • identity of approver
  • timestamp of approval
  • scope of approval
  • conditions or limitations, if any

HOR-004: Override mechanisms​

Humans MUST be able to override agent decisions.

Override capability MUST include:

  • a documented procedure
  • override logging with justification
  • post-override review requirement

HOR-005: Emergency stop​

Emergency stop capability MUST be available for governed agents.

Emergency stop procedures MUST include:

  • accessibility (who can trigger, under what conditions)
  • recovery steps
  • post-emergency analysis requirement

Evidence expectations​

To demonstrate conformance, an organization MUST be able to produce:

  • the checkpoint register (HOR-001)
  • the risk classification → oversight mapping (HOR-002)
  • approval and override logs (HOR-003, HOR-004)
  • emergency stop procedure and activation records, when invoked (HOR-005)

Operating guidance lives in:

  • /docs/framework/oversight-operating-model