Agent architecture advisor

A consultation workflow that encodes the playbook's own decision framework: four questions, then a pattern recommendation — biased toward the playbook's core lesson, "start simple, scale intelligently."

Workflow

1. Ask the four questions, in order, before recommending anything:
   • Control: How much do you need to explain/audit every decision? (high: regulatory, financial, safety-critical / moderate: customer support, content / low: research, brainstorming)
   • Domain complexity: Single well-defined domain, multiple predictable domains, or open-ended/multi-domain?
   • Resources: What's the budget and time-to-market pressure? Multi-agent systems use roughly 10-15x more tokens than a single agent — this should be a real input to the decision, not an afterthought.
   • Domain expertise depth: Does this need deep specialized knowledge in one domain (favors Skills), or coordination across genuinely distinct domains (favors multi-agent)?
2. Map answers to a recommendation using references/pattern-catalog.md — don't skip straight to a sophisticated architecture because it sounds more capable; match complexity to demonstrated need.
3. Default to the simplest pattern that satisfies the control and domain-complexity answers. If a single agent with added Skills could plausibly meet the accuracy bar, recommend trying that before proposing multi-agent coordination.
4. State the upgrade path, not just the current recommendation. Even a single-agent recommendation should note what signal would justify moving to the next tier (see references/decision-framework.md for the e-commerce 5-phase example of this in practice).
5. Name the tradeoff explicitly. Every recommendation should state what's being traded (token cost, latency, engineering time, auditability) — not just what pattern to use.

Composition

Stand-alone advisory skill — typically the first skill consulted before building anything from this repo's patterns/ or agents/ directories.

Quality bar

• All four questions were asked and answered before any recommendation was made.
• The 10-15x token-cost consideration was explicitly weighed, not skipped.
• The recommendation matches the simplest pattern that satisfies the stated constraints — no unjustified complexity.
• An upgrade path (the signal that would justify more complexity later) is stated.