The framework is structured around the seven failure modes called out in the brief, plus the source articles' analysis of why each one happens.
| # | Failure mode | What it looks like | Where this framework addresses it |
|---|---|---|---|
| 1 | Messy | Random employees building random agents on random platforms. Nobody knows what exists. | §6–§8 (CoE + operating model), §15 (registry), §13 (artifacts) |
| 2 | Non-scalable | Each team reinvents intake, prompts, evaluation, tooling, infrastructure. | §9 Pillar 4 (reusable building blocks), §16 (approved stack), §6.3 (CoE evolution to advisory) |
| 3 | Non-maintainable | Agents pile up. Nobody dares retire them. Prompts and tools drift. | §11.2 (per-agent lifecycle), §9 Pillar 4 (versioning + deprecation), §13 (retirement checklist) |
| 4 | Non-secure | Agents have unbounded data access, no scoped credentials, no incident plan. Prompt injection lands. | §17 (privileged identities), §19–§20 (guardrails + control mechanisms), §25 (security), §26 (procurement) |
| 5 | Non-responsible | Bias, hallucination, opaque decisions, no disclosure, no accountability. | §18 (responsible AI), §4 (core principles), §10 (risk classification) |
| 6 | No tracking | No record of who deployed what, what data it sees, what decisions it has made. | §15 (registry), §24 (observability), §13 (artifacts) |
| 7 | No monitoring | Quarterly audits at best; bad outputs run for weeks before anyone notices. | §21 (monitoring signals), §24 (observability), §20 (control mechanisms) |