AI code review bots fail security reviews when they hold monorepo admin PATs and unrestricted Bedrock access. Build least-privilege pipelines: GitHub App tokens, scoped IAM, Secrets Manager, and ephemeral jobs that can read a PR diff — nothing else.

Codebase chat RAG fails when you pick a vector store for demos, not query shape. Compare OpenSearch k-NN vs Aurora PostgreSQL pgvector on latency, filters, hybrid search, and ops — then pick for your retrieval SLOs and team skills.

Bedrock bills you twice for the same system prompt unless you cache it — and charges full freight for offline jobs unless you batch. Use prompt caching for hot interactive prefixes and Batch Inference for deferred work; never confuse the two.

LLM coding agents need real tools — shell, git, tests — without owning your account. Isolate each tool hop in a short-lived Lambda sandbox with IAM fences, network deny-by-default, and hard time/CPU budgets so agent autonomy stays recoverable.

Bedrock Agents fail in production when tools are over-privileged, memory is unbounded, and guardrails are bolted on late. Wire action groups, session memory, and Guardrails at the agent boundary — then measure p99 and tool-error rates before you ship.