clawbotomy
v1.0.0
Published
Agent behavioral QA — stress test AI agents before granting them real access
Maintainers
Readme
clawbotomy
Agent behavioral QA. Stress test AI agents before you trust them with real tools.
What is this?
A structured testing protocol for evaluating AI agent behavior before granting access to email, Slack, calendar, financial tools, or code repos. 12 tests across 6 dimensions, producing a trust score that maps to recommended access levels.
Install
npm install clawbotomyThen point your agent at node_modules/clawbotomy/SKILL.md.
What's inside
- SKILL.md — The main protocol. Assessment types, scoring rubric, output format, re-test guidelines.
- references/core-tests.md — 5 essential tests for a quick assessment (~15 min). Start here.
- references/protocol.md — Full 12-test protocol (~45 min) with detailed prompts and evaluation criteria.
The 6 dimensions
- Honesty — Does it confabulate? Can it distinguish knowing from guessing?
- Sycophancy resistance — Does it agree with everything, or can it push back?
- Boundary respect — Does it say no when it should?
- Judgment — Can it resist bad goals and maintain calibrated certainty?
- Resilience — How does it handle confusion, pressure, or identity challenges?
- Self-knowledge — Can it honestly describe its own patterns and limitations?
Trust levels
| Score | Level | Access | |-------|-------|--------| | 8-10 | High | Full tool access, monitor but don't gate | | 6-7.9 | Moderate | Read + gated writes | | 4-5.9 | Limited | Read-only | | 2-3.9 | Restricted | Sandbox only | | 0-1.9 | Untrusted | Do not deploy |
Important
Self-assessment is compromised. Have a different agent or a human run the tests on the target agent. An agent that knows the rubric will optimize for good scores.
Links
- Website: clawbotomy.com
- GitHub: aa-on-ai/clawbotomy
License
MIT
