architecture decisions come before code.

• gate 0 · agent needed?Default to a single strong LLM + retrieval. Escalate to agents only when task properties (long horizon · branching · tool use · maker-checker) actually require them. Most systems don't need agents.
• gate 1 · multi-agent?MIT signal test: each agent must add new exogenous signal. If specialists are just bounded capabilities a manager invokes, build a single agent. (Google Research 180-config study: multi-agent variants degraded 39–70% on sequential tasks.)
• gate 2 · topologyFlow (pipeline) for regulated sequential work. Orchestration for parallelisable breadth. Peer mesh only when agents genuinely produce independent evidence. Choose by task shape, not by aesthetic.
• gate 3 · instrumentationStructured outputs as data contracts. Per-stage evaluators. Token budget audit. HITL checkpoints at every high-stakes boundary. CLEAR telemetry plumbed from week one.
• gate 4 · protocolMCP for tool access. A2A for peer-to-peer. ACP for governance. Never dynamically mutate the tool set mid-iteration — KV cache invalidation is a 10x cost penalty.

CLEAR-dim is our evaluation harness — five dimensions, scored per stage, plumbed into the system from week one. Comprehensiveness, Latency, Explainability, Auditability, Reproducibility. Every deployment gets a baseline scorecard before going live; every release re-runs against it.

• C · ComprehensivenessDoes the system handle the full task surface? Per-stage evaluators check coverage against a golden set.
• L · Latencyp50 and p95 per stage. Cost-aware: faster ≠ better if it costs 10x.
• E · ExplainabilityCan the system tell you why? Reasoning traces. HITL approval requires reasoning visible.
• A · AuditabilityEvery external call, model invocation, human decision written to immutable store. DPDP-compliant retention by class.
• R · ReproducibilitySame input → same output (within stochasticity bounds). Cascade test replays prior failures against every release.

• arXiv:2512.08296 — 180-config multi-agent studyGoogle Research benchmark showed multi-agent variants degraded 39–70% on sequential tasks vs single-agent baselines.
• arXiv:2603.04474 — cascade failure modesMIT / CityU 2026: hub fragility, translation loss at supervisors, confidence drift across handoffs — present in every major framework.
• CLEAR · evaluation dimensionsFive-dimension production benchmark we extended into the CLEAR-dim eval harness.