Approach A
Rule-based system
Predetermined rules respond to each event.
Risk-aware agents
for private security,
under human oversight
This research investigates how a risk-aware multi-agent AI system can monitor, coordinate, and automate routine private security operations while keeping human supervisors responsible for high-risk decisions.
Proposal · October 2026
ASBOS
A risk-aware multi-agent framework for autonomous private security operations. The design is specified. A prototype has not been built, and no results are reported.
Read the proposal →Main research question
How can a risk-aware multi-agent AI system automate routine private security operations using real-time operational data while preserving human oversight for high-risk decisions?
Design and evaluate a multi-agent framework that can coordinate routine private security operations and change its level of autonomy with operational risk.
- RQ1
Can specialized AI agents coordinate security operations more effectively than a single-agent or traditional rule-based automation system?
- RQ2
How should the level of AI autonomy change according to the operational risk of a decision?
- RQ3
How much human operational workload can be reduced without increasing incorrect or unsafe autonomous actions?
- RQ4
How effectively can a multi-agent system respond to interconnected events involving guard tracking, scheduling, patrol compliance, and incident management?
These questions are unanswered. The comparison that would address them is specified and has not been run.
Architecture
Proposed architecture
Specialist agents watch separate parts of the operation. A coordinator reads their findings and decides whether to act, recommend, ask for approval, or escalate.
Specialist agents
Field Operations Agent
Monitors guard location, geofence status, clock-ins and clock-outs, patrol checkpoints, missed patrols, unexpected movement, and current shift activity.
Workforce Agent
Handles schedules, late arrivals, no-shows, replacement availability, overtime, qualifications, conflicts, and shift coverage.
Incident and Dispatch Agent
Handles incident reports, alarms, requests for assistance, supervisor notification, incomplete reports, follow-up, dispatch, and escalation.
Compliance and Business Operations Agent
Checks licenses, training, site requirements, working-hour limits, timesheet exceptions, overtime policy, and other organizational rules.
Coordinator
Supervisor and Coordinator Agent
Reads the other agents and decides among no action, an automated action, a recommendation, human approval, or immediate escalation.
Workflow
- Monitor
- Detect
- Reason
- Coordinate
- Act or Escalate
High-risk decisions remain under human control.
Agents
Five agent roles
Each role is a responsibility in the design. None of these agents has been implemented.
Field Operations Agent
Monitors guard location, geofence status, clock-ins and clock-outs, patrol checkpoints, missed patrols, unexpected movement, and current shift activity.
Workforce Agent
Handles schedules, late arrivals, no-shows, replacement availability, overtime, qualifications, conflicts, and shift coverage.
Incident and Dispatch Agent
Handles incident reports, alarms, requests for assistance, supervisor notification, incomplete reports, follow-up, dispatch, and escalation.
Compliance and Business Operations Agent
Checks licenses, training, site requirements, working-hour limits, timesheet exceptions, overtime policy, and other organizational rules.
Supervisor and Coordinator Agent
Reads the other agents and decides among no action, an automated action, a recommendation, human approval, or immediate escalation.
Scenarios
Four operational cases
The first prototype is planned around four cases. The sentences below are test requirements, not observed behavior. The full protocols are in the research portal.
Scenario 1
Late arrival or no-show
A scheduled officer does not arrive. The planned test asks the system to detect the gap, judge whether the post is uncovered, consider contact, replacements, qualifications, overtime, and travel, and decide whether a supervisor must approve coverage.
Scenario 2
Guard leaves an assigned geofence
An on-duty officer leaves the assigned site. The planned test asks the system to distinguish normal activity, an authorized exception, a temporary problem, a coverage gap, or an event that needs a supervisor. A location anomaly is not to be treated as employee misconduct.
Scenario 3
Missed patrol checkpoint
A required checkpoint is not completed on time. The planned test asks the system to consider recent checkpoints, elapsed time, location, site rules, active incidents, and officer status before it continues monitoring, contacts the officer, recommends an action, or escalates.
Scenario 4
Incident with coverage impact
An incident pulls an officer off a post or requires more people. The planned test asks the field, workforce, and dispatch agents to coordinate response, replacement coverage, supervisor notification, documentation, and follow-up.
Autonomy
Levels 0–4
ASBOS adapts levels of human interaction with automation (Parasuraman, Sheridan, and Wickens, 2000) to private security operations. High-risk decisions stay with a person at every level above observation.
Level 0
Observation
The system monitors operations, identifies events, and records information. No operational decision is made automatically.
Level 1
Recommendation
The system detects a problem, analyzes the situation, and recommends an action. A human makes the final decision.
Level 2
Approval required
The system determines and prepares an action. A human supervisor must approve it before it is carried out.
Level 3
Routine autonomy
The system may carry out predefined low-risk operational actions. The action is logged and available for human review.
Level 4
Coordinated autonomy
Specialist agents may resolve an event together when it stays inside established policy and risk boundaries. High-risk decisions remain under human control.
Methodology
Three approaches
An event-driven prototype will give the same simulated, or de-identified, scenarios to three systems. This comparison is planned. It has not been run. The method is specified in the research portal.
Approach B
Single AI agent
One agent receives the operational information and chooses a response.
Approach C
ASBOS multi-agent system
Specialist agents analyze their own areas and coordinate through the proposed architecture.
A system that performs more autonomous actions but produces unsafe or incorrect decisions will not be considered superior.
- Successful event-resolution rate
- Average response time
- Incorrect action rate
- Unsafe action rate
- Correct escalation rate
- Unnecessary escalation rate
- Number of human interventions
- Percentage of operational steps automated
- Decision consistency
- Policy violations
- Decision auditability
- Estimated reduction in human operational workload
The proposal names these measures. It does not yet define how they will be scored.
Status
Where the work stands
October 2026. The public record is a proposal, not a results paper. Nothing here should be read as evidence that the multi-agent design performs better than the alternatives.
| Work | State |
|---|---|
| Problem statement, research gap, and research questions | Specified |
| Five-agent architecture and coordination workflow | Specified |
| Autonomy levels 0–4 | Specified |
| Four operational scenarios | Specified |
| Comparative method and named evaluation measures | Specified |
| Research budget estimate | Specified |
| Event-driven prototype | Planned |
| Simulated or de-identified operational environment | Planned |
| Rule-based, single-agent, and multi-agent runs | Planned |
| Scoring rules for the evaluation measures | Not yet defined |
| Measured results | Not available |
Ethics
Human oversight
Guard tracking and workforce monitoring raise privacy and employment questions. Location should be collected only when the operation needs it, and a location anomaly should not be read as evidence of misconduct. Recommendations and automatic actions are to be logged so a supervisor can see what information was considered and why an action was proposed or taken.
Defined human responsibility matters when automated systems are used in operations (Tabassi, 2023). Recent work on agentic systems also argues for oversight through the workflow, not only after the fact (Dhanorkar, Passi, and Vorvoreanu, 2026).
The research prototype will not be allowed to independently make high-impact decisions:
- Employee termination
- Disciplinary action
- Hiring decisions
- Law-enforcement actions
- Accusations of misconduct
- Other significant employment decisions
The study is about operational coordination. It is not a proposal for automated punishment.
People
Researcher and advisor
Researcher
K.M. Fazle Rabbi
Department of Computer Science & Engineering
St. Cloud State University
Research advisor
Dr. Indira Dutta
Proposal dated October 2026.
Research portal
The proposal, methodology, experiment plan, references, and downloads are published separately at research.asbos.org.
Open the research portal