Verification Agent
A read-only adversarial agent that continuously audits other agents' actions.
The /verify skill acts on a specific deliverable at a point in time. The verification agent goes further: it's a permanent agent that observes and audits all other agents' actions in real time.
The verification agent has no write permissions. It cannot modify files, execute commands, or send messages. Its sole role is to observe, analyze, and alert.
This separation of roles is fundamental for security and reliability. An independent, read-only agent that observes from the outside is far more reliable than an agent that checks itself.
Passive observation
The verification agent receives a real-time stream of all actions executed by other agents.
Continuous analysis
Each action is evaluated against a set of rules: permission compliance, consistency with instructions, adherence to best practices.
Anomaly detection
The system identifies suspicious patterns: unusual action repetition, access to unexpected resources, inconsistent results.
Alerts
Detected anomalies generate alerts classified by severity: info, warning, critical. Critical alerts interrupt the current action.
Reports
Periodic audit reports are generated, summarizing observations, detected anomalies, and recommendations.
Enhanced security
The verification agent detects that an agent is trying to access a file outside its authorized scope. The action is blocked and an alert is sent.
Code quality
The verification agent observes that an agent commits code without tests. It generates a warning and recommends adding tests before merging.
Compliance
In a regulated environment, the verification agent ensures all actions comply with defined policies (data retention, sensitive information access).
featurePages.verificationAgent.config
- Available for all multi-agent EasyClaw v2 installations
- Disabled by default, recommended for production environments
- Requires at least one other active agent to monitor