rag-vs-agent-architectlisted

# RAG vs Agent Architect Architecture follows problem shape. The recommendation isn't credible until it says how each rejected option would fail *here* — and how the chosen one fails too. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Rejected failure modes stated:** every rejected architecture gets its failure mode for THIS problem — what breaks, wastes, or risks given the stated shape. Generic cons ("agents are complex") and benefits-only recommendations fail the gate. - **G2 — Chosen option's failure mode + mitigation:** the recommended architecture carries its own dominant failure mode and the concrete mitigation. An architecture presented as failure-free fails. - **G3 — Shape-derived, input-only:** the recommendation cites the problem properties it turns on, quoted from the input (actions or answer-only, corpus dynamics, stakes, volume, latency). No invented constraints, budgets, or compliance rules. ## The shape rubric | Property | Pulls toward | |---|---| | Answer-only, no actions | RAG | | Known, enumerable operations (lookup, file, update) | tool-calling | | Open-ended multi-step work, plan varies per request | agent | | Distinct sub-problems with different shapes | hybrid (composed, not defaulted) | | Static/versioned corpus | RAG index | | Live/system-of-record data | tools over retrieval | | High stakes per answer | fewer moving parts + human/citation checks | | Tight latency budget | fewer hops — agents pay per step | ## St