Name the owner
Who is accountable for this work and the decision that follows?
Sophizo research / Operating note 01
Who checks the agents?
AI agent oversight research on human review, agent span of control and safe handoffs.
Read the findings
The question is not whether an agent can produce an output. It is whether a named person has enough context and capacity to review the action, including the work passed between agents.
This report brings together two practical ideas: review capacity is shaped by the work and its risk, and authorization must remain intact when agents hand work to one another. It does not claim a single safe ratio or present market-wide measurement.
The findings
Four conditions for review that holds.
- 01
The review burden is an operating constraint.
Adding agents increases the amount of work a person may be expected to review. Capacity depends on role, task complexity and risk, not a universal agent-to-person ratio.
- 02
Risk changes the review rule.
Reading and drafting are not the same as changing a record or making a consequential decision. Oversight should track the action and its possible impact.
- 03
Handoffs can break the chain of accountability.
A safe agent can still create an unsafe outcome when the next step loses the human origin, expands permission or trusts an incomplete record.
- 04
Governance must travel with the action.
A policy written at setup is not enough. Review the authority, evidence and owner at each consequential action and each delegation boundary.
Exhibit 01 / Span of control
Agent count alone does not tell you review capacity.
The Ratio Problem describes Agent Span of Control as a role-specific calculation: base task capacity divided by risk weight times complexity factor. The inputs are a way to make assumptions visible, not a universal benchmark.
Base task capacity ÷ (risk weight × complexity factor)
Passive work that reads and drafts may be reviewed differently from active work that changes records or communicates externally. The responsible operator should account for the actual work, exposure and judgment required.
Exhibit 02 / Risk-linked review
Set the review rule before the action.
R1 to R4 is a practical framework for matching oversight to risk. It makes a distinction between low-impact observation and actions that need a person to approve first.
- R1
Read and draft
Read information and prepare drafts. Log the work.
Sample the work - R2
Change internally
Make internal changes within the signed rulebook.
Review exceptions - R3
Act externally with approval
A person approves the action before it reaches anyone outside the business.
A person approves before external action - R4
Prepare only
A person decides and executes. The agent prepares the work only.
A person decides and executes
Risk sets the review rule. The person accountable still sets the boundary.
Exhibit 03 / The handoff
Safe agents still need safe handoffs.
When an agent delegates, check whether the next step preserves origin and attribution, narrows authority, respects the human ceiling, limits access lifetime, checks authorization at runtime and leaves evidence.
- 01Origin
Can the receiving agent prove who authorized the work?
- 02Attribution
Does the record name every agent that acted?
- 03Narrowing
Is delegated capability smaller than the authority upstream?
- 04Human ceiling
Is effective authority capped by what the originating person could do?
- 05Lifetime
Is access limited to this task, resource and period?
- 06Runtime check
Is authorization checked when the consequential action happens?
- 07Evidence
Can the chain be reconstructed and a link revoked?
Framework adapted from Sophizo's Safe Agents, Unsafe Handoffs.
A practical review loop
Ask these questions at the boundary.
Set capability
What can the agent read, prepare, change or send?
Match oversight
What is sampled, checked, approved or stopped?
Keep evidence
Can a reviewer reconstruct the action and its handoffs?
Keep the framework
Get the report as a PDF.
Request the printable version of Who checks the agents?
Scope and sources
What this report is based on.
This report synthesizes only two Sophizo essays: The Ratio Problem, which frames role-specific review capacity, and Why safe agents can still create unsafe handoffs, which outlines the seven handoff checks. The R1 to R4 review ladder is presented here as an operating framework, not external research.
No market adoption rates, client outcomes or vendor claims are used as evidence in this report.