Pattern Selection Map¶

Compare patterns by what they cost, where they break, and what they assume — then pick the cheapest one that solves your problem, not the most sophisticated.

Adopting agent patterns without comparing their costs leads to two failure modes documented on this site: stacking sophisticated patterns (cargo-cult agent setup) and stacking frontier-model roles until economics collapse (compound engineering's 80/20 inversion). The matrix below surfaces the trade-offs already documented on each pattern's canonical page so you can compare across them without re-reading every one.

The matrix is scoped to this site's 14 patterns that map onto a common set of axes. Patterns not on the matrix exist for cases where these axes are not the dominant trade-off — they live on their own pages.

The Matrix¶

Pattern	Token cost	Latency overhead	Frontier-model dependency	Blast radius	Verification cost	Task class
Harness Engineering	low	none	none	contained writes	linter-able	open-ended
Agent Self-Review Loop	high	+N turns	one role	contained writes	eval-able	iterative
Cognitive Reasoning vs Execution	medium	+1 turn	one role	contained writes	linter-able	iterative
Episodic Memory Retrieval	low	+1 turn	none	read-only	eval-able	iterative
Agent Circuit Breaker	low	none	none	read-only	linter-able	iterative
Agent Backpressure	low	none	none	contained writes	linter-able	iterative
Multi-Agent Topology Taxonomy	medium	+N turns	one role	contained writes	human-only	open-ended
Declarative Multi-Agent Topology	medium	+N turns	one role	contained writes	linter-able	open-ended
Economic Value Signaling	medium	+N turns	none	production effects	human-only	open-ended
Review-Then-Implement Loop	high	+N turns	one role	contained writes	eval-able	iterative
Agentic Code Review Architecture	high	+N turns	one role	read-only	eval-able	iterative
Incremental Verification	low	+1 turn	none	contained writes	linter-able	iterative
Escape Hatches	low	none	none	read-only	human-only	open-ended
Compound Engineering	very high	unbounded	all roles	production effects	human-only	open-ended

Axis Legend¶

Token cost — relative to a baseline single-shot prompt against the same task:

low — fixed overhead or one-time setup; per-task addition is negligible
medium — roughly 1.5× to 2× baseline
high — roughly 2× to 5× baseline (multiple review or critic passes per task)
very high — 5×+ baseline or stacks frontier-model roles across plan/work/assess phases

Latency overhead — wall-clock impact on time-to-result:

none — runs in parallel or as setup; no per-task added latency
+1 turn — adds one model round-trip
+N turns — adds a bounded loop (typically 2–5 iterations or fan-out turns)
unbounded — loops until convergence or human intervention

Frontier-model dependency — how many roles require a top-tier model to function:

none — works with mid-tier or open-weight models
one role — one role (reasoning, reviewer, or planner) benefits materially from a frontier model; the rest can run on cheaper models
all roles — every role (planner, executor, critic, reviewer, summariser, memory) benefits from a frontier model; the most expensive shape

Blast radius — the maximum reach of a failure or unintended action:

read-only — the pattern only observes or gates; no writes outside its own state
contained writes — writes are scoped to feature branches, ephemeral state, or pre-approved paths
production effects — writes reach production systems, shared ledgers, or persistent infrastructure without an automatic gate

Verification cost — how the pattern's correctness is checked:

linter-able — a deterministic script or type check confirms the pattern is wired correctly
eval-able — needs an LLM-graded or metric-based eval to confirm quality; deterministic checks are insufficient
human-only — outcomes require human judgement to validate (taste, architecture review, market dynamics)

Task class — the kind of work the pattern fits:

one-shot — single-prompt, single-result tasks
iterative — bounded loops with a clear convergence criterion
open-ended — long-running work without a fixed convergence point (greenfield, ongoing maintenance, multi-agent operations)

Why It Works¶

The matrix compresses information that is already present and sourced on each pattern's canonical page. Experienced engineers already think in trade-off axes when choosing architectural patterns — they just lack a centralised comparison surface. Cognitive offloading of the cross-page comparison step is the mechanism. The axes were chosen because they are the dimensions where stacking patterns blindly produces the documented failure modes: token economics, latency budgets, frontier-model cost, and blast radius are the four levers that the 80% problem in agentic coding traces production failures back to.

When This Backfires¶

The matrix is a comparison aid, not a recommendation engine. Three failure modes to watch:

Pattern shopping — scanning the table and assembling several patterns at once produces the exact stack-everything failure mode the page exists to defuse. The TL;DR and the closing rule are deliberate counterweights.
Stale rows — pattern pages evolve over time. If a pattern page changes its cost or blast-radius characterisation, the matrix row diverges silently until the next periodic audit catches it. The last_reviewed frontmatter dates the synthesis.
Axis flattening — a single ordinal value per axis hides distributions. A pattern marked medium token cost in steady state may spike to high during cold start or on certain task shapes. The canonical page carries the nuance; the matrix row does not.

Key Takeaways¶

Compare across patterns on costs you already care about — token spend, latency, blast radius — before composing them.
low-cost patterns (harness engineering, agent backpressure, incremental verification) are the unsexy foundation; sophisticated patterns assume these are already in place.
very high cost patterns (compound engineering) deliver on long horizons but collapse economics on short tasks — match cost class to task class.
production effects blast radius (economic value signaling, compound engineering) requires human-only verification — these are not patterns to stack speculatively.
If two patterns score similarly across all axes, pick the one your team can debug at 3am.

Patterns — the parent index this map sits under
Cargo Cult Agent Setup — copying patterns without understanding the trade-offs is the failure mode this map exists to defuse
AI Development Maturity Model — adoption phases for AI coding tools; complements the per-pattern trade-off view with a developmental view
Cost-Aware Agent Design — the broader treatment of cost as a first-class design constraint
Compound Engineering — the highest-cost pattern on the matrix; the planning/review investment that makes it economic