You are accountable for
- Controls as platform defaults, not per-team homework
- Prompts, configs, and evals in the same lifecycle as code
- Clear guardrail APIs: screening, grounding, identity, quotas
What keeps you up at night
- An indirect injection turning retrieved content into commands
- A prompt tweak shipping a behavior change no eval caught
- Being handed 'comply with the policy' with no enforcement primitives
Decisions you own
- Structure the pipeline: screen input, ground generation, validate output
- Wire eval gates into CI for every prompt/model/retrieval change
- Scope every credential — app, tool, agent — to least privilege
Your layers of the stack
Questions you should be asking
- What does the platform give me for free: screening, logging, evals, identity?
- Where are the approved patterns for RAG, agents, and tool use?
- How do I test against injection and jailbreak before launch?