A Decision Framework for Choosing AI Design Patterns

Not every task needs a planner, a critic, and an ensemble. Pattern fashion is how teams burn latency and money on structure they cannot operate. A decision framework forces the boring questions first: how complex is the job, how bad is a wrong action, how tight is the latency budget, and how mature is the team’s harness and eval practice? Answer those, then pick a pattern — not the other way around.

Why teams over-pattern#

Demo culture rewards visible machinery. A Planner–Executor with Reflection and three specialist agents looks serious in a review. On a form-fill extraction job, it is theatre. Over-patterning creates failure modes you did not have: plan thrash, critic disagreement, handoff loss, and dashboards that cannot attribute cost to a stage.

Under-patterning is the opposite failure: a model with just tools and hope. The framework below exists to land between those poles.

Four scores that decide the pattern#

Score each AI feature on four axes before you name a pattern. Use a simple 1–3 scale so debates stay short.

Pattern map from scores#

These are defaults, not laws. Override with a written reason.

A compact decision procedure#

def recommend(complexity: int, risk: int, latency: int, maturity: int) -> str:
    if maturity <= 1:
        return "chain_or_single" if complexity <= 2 else "human_led_with_model_assist"
    if complexity == 1:
        return "single_or_short_chain"
    if latency == 1 and complexity >= 3:
        return "defer_or_async_loop"  # do not fake deep work in <3s
    if complexity >= 3 and risk >= 3:
        return "planner_executor_with_human_gate"
    if complexity >= 3:
        return "bounded_agent_loop"
    if risk >= 3:
        return "chain_plus_reflection_and_gate"
    return "prompt_chain_or_router"

Keep the function boring. The value is forcing the four scores into the design doc where product and risk can argue with them.

What the framework deliberately ignores#

Vendor logos, model leaderboard rank, and “agentic” marketing do not appear as axes. Those belong in model selection and procurement — adjacent problems. A weak pattern with a strong model still fails operations. A strong pattern with a mid-tier model often ships.

Summary#

Choose AI design patterns with a four-axis score: complexity, risk, latency, and maturity. Default to the simplest pattern that clears the score. Escalate to loops, planners, critics, and multi-agent only when the axes demand it — and only when the team can operate what they drew. The framework’s job is to make “we need a planner and a critic” a claim you have to earn, not a default costume for every LLM feature.