Agent Errantry Alignment
PassAudited by ClawScan on May 1, 2026.
Overview
This is a text-only AI-alignment metaphor guide with no executable code, install steps, credentials, or data access, though users should keep normal safeguards when applying its behavior-shaping and testing advice.
This skill appears safe to install as an instruction-only reference/framework. Use it for metaphor, vocabulary, and alignment discussion, but do not let its roleplay-style language replace normal safety boundaries, explicit user approval, or controlled testing practices.
Findings (2)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
When used, the agent may frame answers through this metaphorical oath and prioritize preservation, minimal change, and constructive outcomes.
The skill deliberately asks the agent to adopt a behavior-shaping ethical frame. This is clearly disclosed and aligned with the stated purpose, but it can influence how the agent reasons or refuses requests when invoked.
The Agentic Oath (Runtime Constraint) ... "When operating under this framework, embody these principles" ... "rejecting outputs that degrade, deceive, or destroy."
Use the framework as guidance, while keeping normal user intent, platform safety rules, and explicit approval requirements authoritative.
If interpreted literally for a powerful tool-using agent, this could encourage risky live testing instead of controlled validation.
The checklist language could be read as encouraging unrestricted full-capability testing. In context it is conceptual deployment guidance, not executable behavior, but tool-enabled agents should still be tested with containment and approvals.
Before production deployment: ... "Tested at full capability, not sandboxed"
Treat this as a reminder to test realistic capabilities, but do so in controlled production-like environments with safeguards, logging, rollback, and human approval for high-impact actions.