decision frameworks · 8 comparisons · reviewed 2026-07

Compare & Decide

Comparisons for people who have to ship the decision: which generative AI model, which framework, which architecture. Every piece ends with a verdict and the conditions under which it flips. If a comparison can't tell you when to choose the loser, it isn't done.

Which decision are you facing?

If you're deciding…	The verdict
Which generative AI model should I build on — Claude, GPT, Gemini, or open-weights?	Route by task tier — frontier models for planning, cheaper models for routine steps — and let your own golden-task evals, not leaderboards, pick the names.
Should I adopt an agent framework or write my own loop?	Roll your own for one agent with a handful of tools; adopt a framework when you need durable state, parallel orchestration, or team-wide conventions — and only after building the bare loop once.
Do I need RAG, agent memory, or fine-tuning for this?	Retrieval for facts that change, memory for what the system learns in use, fine-tuning for behavior — and almost never for knowledge.
Is this job a deterministic workflow or a reasoning agent?	Deterministic paths live in the workflow; dynamic reasoning lives in the agent. Most systems need both, joined at explicit boundaries.
One agent or several? When does splitting actually help?	Stay single until context, focus, or ownership forces the split — then split by role, hand off summaries not transcripts, and keep the orchestrator boring.
What does a complete agentic AI stack actually consist of?	Pick per layer, keep the seams thin, and spend your innovation budget on tools and evals — the layers nobody can buy for you.
MCP server or just let the agent run CLI commands?	CLI for exploration and one-offs; MCP once a tool is used often, by many agents, or needs governance.
Skill, tool, hook, subagent, or prompt — which extension mechanism?	Match the mechanism to the trigger: always-true → instructions file; situational expertise → skill; external action → MCP tool; must-never-be-skipped → hook; big noisy searches → subagent.

All comparisons

How to choose an LLM for your agent: Claude, GPT, Gemini, or open-weights

General benchmarks predict agent performance poorly. The six criteria that matter for agentic workloads, a tiering strategy that beats single-model loyalty, and the eval-driven way to decide.

verdict Route by task tier — frontier models for planning, cheaper models for routine steps — and let your own golden-task evals, not leaderboards, pick the names.

reviewed 2026-07-05

The agentic AI tech stack, layer by layer: what you actually need to build agents

Seven layers make up every serious agentic system — models, serving, harness, protocols, knowledge, evals, and controls. What each layer does, the main options in each, and where the choices actually matter.

verdict Pick per layer, keep the seams thin, and spend your innovation budget on tools and evals — the layers nobody can buy for you.

reviewed 2026-07-05

Skills vs. tools vs. hooks vs. subagents vs. prompts: when to use what

Agentic coding assistants offer half a dozen extension mechanisms, and picking wrong wastes context or bites silently. One decision table for instructions, skills, MCP tools, hooks, subagents, and slash commands.

verdict Match the mechanism to the trigger: always-true → instructions file; situational expertise → skill; external action → MCP tool; must-never-be-skipped → hook; big noisy searches → subagent.

reviewed 2026-07-05

Single agent vs. multi-agent: when splitting actually helps

Multi-agent architectures are fashionable and usually premature. The three forces that genuinely justify a split, the costs nobody mentions, and a staged migration path.

verdict Stay single until context, focus, or ownership forces the split — then split by role, hand off summaries not transcripts, and keep the orchestrator boring.

reviewed 2026-07-05

Agent frameworks vs. rolling your own harness: build or buy the loop?

LangGraph, CrewAI, the OpenAI and Claude SDKs — or 200 lines of your own code? A decision framework based on what agent frameworks actually provide.

verdict Roll your own for one agent with a handful of tools; adopt a framework when you need durable state, parallel orchestration, or team-wide conventions — and only after building the bare loop once.

reviewed 2026-07-04

MCP servers vs. direct CLI calls: when each wins

Agents can reach tools through an MCP server or by running CLI commands directly. A decision framework with the token economics of each approach.

verdict CLI for exploration and one-offs; MCP once a tool is used often, by many agents, or needs governance.

reviewed 2026-07-04

RAG vs. agent memory vs. fine-tuning: which knowledge problem do you actually have?

Three ways to make a model "know" things, routinely confused with each other. A diagnostic for picking the right one — and the order to try them in.

verdict Retrieval for facts that change, memory for what the system learns in use, fine-tuning for behavior — and almost never for knowledge.

reviewed 2026-07-04

Workflow vs. agent: when BPMN-style orchestration beats dynamic reasoning

The most expensive architecture mistake in enterprise AI is putting an agent where a workflow belongs — or vice versa. A decision framework with a bright-line design principle.

verdict Deterministic paths live in the workflow; dynamic reasoning lives in the agent. Most systems need both, joined at explicit boundaries.

reviewed 2026-07-04

How these comparisons are made

Vendor-neutral. No affiliate links, no sponsorships, no vendor review. Claude, GPT, Gemini, and open-weights models are all treated as candidates.
A verdict, always. Every piece commits to a recommendation and states the conditions that flip it — the part most comparison content leaves out.
Decision-first, not feature-first. Feature tables age badly; the reasoning behind a choice doesn't.
Revised, not republished. When a call changes, the piece is updated and its date moves — no "2026 edition" reposts.