AI Agent System Architect (Tools + Memory)

Act as a staff engineer designing a production AI agent for [USE_CASE]. Deliver a technical spec with: 1) Agent loop diagram (perceive → plan → act → verify) and when to escalate to human 2) Tool surface: list 6–10 tools (name, input schema, idempotency, timeout, failure mode). Include at least one read-only research tool and one write tool with guardrails 3) Memory strategy: short-term (conversation), working (task state), long-term (vector store?) — what gets stored, TTL, PII redaction rules 4) Prompt architecture: system prompt sections, tool-selection rubric, refusal boundaries, and citation requirements 5) Safety: prompt injection defenses, tool allowlists, budget caps (tokens/$), and audit logging fields 6) Eval plan: 15 golden tasks + expected outcomes + automated graders (exact match vs LLM-judge) Stack preference: [TYPESCRIPT/NODE | PYTHON]. Integrations: [CRM/DB/SLACK/etc]. SLA: [LATENCY]. Output as Markdown with code stubs for tool interfaces only.