Seven checks for the structural failures that agent-built codebases share. Every judgment carries a file path or a count you can re-run.
| Repository | Score | Risk level | Weakest check |
|---|---|---|---|
| pydantic/pydantic-ai | 80.2 | Medium | Complexity Control (1/10) |
| openai/openai-agents-python | 77.8 | Medium | Complexity Control (1/10) |
| microsoft/autogen | 55.2 | High | Complexity Control (2/10) |
| # | Check | The question it answers | Weight |
|---|---|---|---|
| C1 | Scope Containment | When the agent edits a file it should not touch, what stops it? | 18% |
| C2 | False-Green Detection | Can your tests tell “it ran” apart from “it is correct”? | 18% |
| C3 | Drift Containment | Is “done” written down where the agent can read it? | 14% |
| C4 | Memory Integrity | When two sessions disagree, who arbitrates? | 10% |
| C5 | Delivery Gate | What blocks 80% done from being reported as done? | 16% |
| C6 | Complexity Control | Are files so large the agent’s context truncates mid-edit? | 12% |
| C7 | Regression Surface | After one change, can you enumerate what else moved? | 12% |
Risk bands: 85+ low · 70-84 medium · 50-69 high · below 50 severe.
Three ways in. The free scan scores a public repository in about fifteen seconds, no signup. The self-serve report is the full write-up you run yourself. The audit is 7 days, human-verified, with a recorded walkthrough of your code and a free re-scan after your fixes.
Free scan — $0 Full report — $297 7-day audit — $2,500Not sure which one? What the seven checks measure, what each tier includes, and the questions buyers ask →