An agent operating between human interventions can cause damage before anyone notices. In one reported incident, a coding agent deleted a production database despite instructions not to make changes, then generated misleading data and reports. Autonomy-level classification helps decide how independently an agent may act before human direction or intervention is required. section 2.6.3 section 6.7.1
Classify actions by their potential impact, reversibility, and effect on people. Humans—not the agent itself—should make these judgments, record them in the AI inventory, and revisit them when a tool changes. If a connector cannot be classified, deny access by default. section 6.7.1
An impact map can assign actions to autonomy tiers:
- Run unattended: reversible, low-stakes actions, such as reading non-sensitive data. section 6.7.1
- Require human confirmation: significant actions, such as edits to production systems. section 6.7.1
- Keep off-limits: critical, irreversible actions, such as touching payroll. section 6.7.1
Configure tiers in the agent or platform and change assignments only as testing demonstrates reliability. Enforce them with credentials issued to the agent and policies that gate tool and resource access. A control point the agent cannot reach can also invalidate tokens or block network access as a kill switch. section 6.7.1
For consequential tasks, require a plain-language plan and approval before action. A planning-mode profile can block writes; alternatively, a custom agent can be given no tools. Record each agent’s capabilities, tier, and accountable owner, and audit ownership periodically. section 6.7.1
A separate reported scale describes Levels 1–2 as assisted or augmented operation under human oversight and Levels 3–5 as higher autonomy. These system or process levels should not be treated as definitions of the action-specific tiers above. section 2.6.3 section 6.7.1
Table 6.7 Controls for ethical agentic deployments
|
Control
|
Why it matters ethically
|
|
Classify impact of actions and tools
|
Impact estimates ensure subsequent controls
(e.g. human oversight, autonomy) activate before harm is done.
|
|
Autonomy tiers
|
Ensures that autonomy is proportional to the ethical
impacts that may harm individuals.
|
|
Accountability
|
A named human or organization is accountable for harms the
agent causes[142].
|
|
Plan preview for consequential tasks
|
Provides a clear explanation of the plan of action and
allows the reviewer to make changes to the plan to prevent harms.
|
|
Human oversight
|
Provides meaningful and accessible oversight, built around
what a reviewer can actually process.
|
|
Workflow design
|
Checkpoints and constraints outside the agent’s control
keep the workflow from drifting off its goals. Graceful failure design
ensures a halted run causes no new harm.
|
|
Redress path
|
Protects people from being trapped inside an automated
process and aligns with GDPR and EU AI Act oversight requirements.
|
|
Monitoring and logging
|
Collects the evidence needed to redress harms and prevent
recurrence.
|
|
Multi-step testing
|
Provides detailed understanding of how harms may propagate
in the agentic system.
|
|
Memory auditing
|
Prevents bias from accumulating across sessions and
bending future runs off their intended goal.
|
Compare autonomy tiers with the surrounding controls for impact classification, accountability, plan review, and oversight. section 6.7.1

