One-Page Cheat Sheet

Use this page when you need the fastest possible answer to: which tool belongs to which layer, and what should it produce?

The 30-second model

flowchart TB
    A[Model API or local LLM] --> B[Agent app framework]
    B --> C[Agent harness/runtime]
    C --> D[Workflow/methodology]
    D --> E[Artifacts and source of truth]
    E --> F[Repo, CI, evals, audit]

    B --> B1[LangChain / LangGraph]
    C --> C1[Codex CLI / Claude Code / Hermes]
    D --> D1[Spec Kit / OpenSpec / AI-DLC / GSD / Superpowers]

The key idea:

Most frameworks share verbs like plan, implement, and review. They differ by what they govern.

Choose by problem

Your real problem	Best starting point	Why
The agent guesses requirements	GitHub Spec Kit	Make intent, spec, plan, and tasks explicit before coding
You want lightweight SDD	OpenSpec	Keep change proposals and delta specs without heavy process
You need enterprise traceability	AWS AI-DLC Workflows	Add approvals, risk records, NFRs, and audit trail
Work spans many sessions	GSD	Preserve context, phase plans, and handoffs
Agent coding lacks discipline	Superpowers	Enforce design, TDD, review, and finish habits
You are building a RAG or tool-calling app	LangChain	Compose models, prompts, tools, retrievers, and chains
You are building a stateful agent service	LangGraph	Model state, nodes, edges, checkpoints, and human-in-the-loop
You need a customizable agent runtime	Hermes	Own the harness, tools, memory, skills, and model routing
You need a protocol for tools	MCP	Standardize tool exposure and integration boundaries

Same words, different ownership

Word	Workflow framework meaning	Harness/runtime meaning	App framework meaning
Plan	Delivery plan, spec, tasks, approval path	Agent execution plan for tool use	Graph, chain, node, or state transition design
Implement	Code changes against requirements	Tool calls, file edits, terminal actions	Runtime logic inside an AI app
Review	Human/code/spec review	Agent output verification	Eval, trace, or behavior inspection
Memory	Project context or long-lived handoff	Agent memory/session context	Application memory, state, or retrieval
Governance	Delivery gates and accountability	Tool permissions and sandboxing	Runtime guardrails and eval thresholds

Selection matrix

Context	Primary workflow	Supporting layers	Minimum artifacts
Small product feature	Spec Kit or OpenSpec	Superpowers	spec/change proposal, tasks, tests
Startup MVP	OpenSpec	Superpowers, LangChain if AI app	change proposal, test checklist, done criteria
Enterprise modernization	AWS AI-DLC	Spec Kit, Superpowers	risk record, NFRs, approval log, migration plan
RAG product	OpenSpec	LangChain, evals, observability	data contract, eval set, prompt contract
Long-running agent service	AI-DLC or OpenSpec	LangGraph, security/governance	state schema, tool policy, eval gates
Internal agent platform	AI-DLC	Hermes, MCP, LangGraph, observability	tool registry, memory policy, audit trail
Multi-agent delivery project	GSD	Superpowers, Spec Kit/OpenSpec	phase plan, context packet, handoff notes

Do not compare these directly

Wrong comparison	Better framing
LangGraph vs AI-DLC	LangGraph builds runtime behavior; AI-DLC governs delivery
Hermes vs Spec Kit	Hermes runs agents; Spec Kit structures specs
MCP vs LangChain	MCP exposes tools; LangChain builds app logic that may use tools
Superpowers vs Codex CLI	Superpowers is engineering method; Codex CLI is an agent harness
OpenSpec vs LangGraph	OpenSpec governs changes; LangGraph implements stateful agent behavior

Minimal stack by maturity

flowchart LR
    A[Prototype] --> B[OpenSpec + Superpowers]
    B --> C[Production AI app]
    C --> D[OpenSpec + LangChain + evals]
    C --> E[LangGraph + tool policy]
    E --> F[Enterprise agent system]
    F --> G[AI-DLC + LangGraph + MCP + observability + audit]

Rule of thumb

Use one primary workflow as the source of truth. Add app frameworks, harnesses, tools, and eval layers around it, but do not let every layer invent its own plan.