references/architecture-selection.md

# Architecture and autonomy selection
A useful agent system starts with a product goal and an explicit risk budget, not a framework. Ascending levels: (0) deterministic function, (1) bounded LLM call, (2) tool-enabled LLM, (3) deterministic multi-step workflow, (4) bounded autonomous loop, (5) multi-agent delegation. Try the smallest sufficient level first.

Compare each candidate across measurable success, variable decision count, need for external data, failure recovery, latency, price, debuggability, and user trust. Rejected candidates should have a reason. Sample: triage → classify ticket with structured outputs → retrieve account data with a read tool → draft response → human approves send. A five-agent team is unnecessary without separate tool permissions, scale, or responsibilities.

Do not imply a workflow lacks nondeterminism because a model call is inside it. Define retries, stop rules, output validation, side effects and fallback even for simple workflows.

## Decision record
Problem and user; accepted input; safety risks; alternatives considered; chosen level and justification; quality/cost targets; failure modes; disconfirming benchmark; escalation condition. Use `templates/architecture-decision.md`.