Agent Harness vs Workflow Framework

The most common confusion in AI engineering is mixing up where the agent runs with how the agent should work.

For the larger architecture context, read the AI Engineering Stack Map first. Harness vs workflow is only one distinction inside the full stack.

flowchart TB
    A[LLM / Model] --> B[Agent App / Orchestration Framework]
    B --> C[Agent Harness / Runtime]
    C --> D[Workflow / Methodology]
    D --> E[Artifacts and Source of Truth]
    E --> F[Repo / CI / Deployment]

    B1[LangChain] --> B
    B2[LangGraph] --> B
    C1[Hermes] --> C
    C2[Codex CLI] --> C
    C3[Claude Code] --> C
    D1[Spec Kit] --> D
    D2[OpenSpec] --> D
    D3[AI-DLC] --> D
    D4[GSD] --> D
    D5[Superpowers] --> D

The layers

Layer	Examples	It answers
Model	GPT, Claude, Hermes models, local LLMs	What reasoning engine generates responses?
Agent app/orchestration framework	LangChain, LangGraph, LlamaIndex, Semantic Kernel	How do I build an AI app or stateful agent system?
Agent harness/runtime	Codex CLI, Claude Code, Hermes Agent, OpenCode, Cursor Agent	Where does the agent run and how does it use tools?
Workflow/methodology	Spec Kit, OpenSpec, AI-DLC, GSD, Superpowers	What process should the agent follow?
Artifact/source-of-truth	specs/, openspec/, aidlc-docs/, .planning/, tests	What must the agent obey?
Repo/CI/deployment	Git, test runner, CI, GitHub Pages, production	Where does evidence and delivery happen?

Why this distinction matters

If you confuse the layers, you ask the wrong comparison:

Wrong: Hermes vs OpenSpec
Better: Hermes + OpenSpec

Wrong: Codex CLI vs Spec Kit
Better: Codex CLI running a Spec Kit workflow

LangChain and LangGraph are app/orchestration frameworks. Hermes, Codex CLI, and Claude Code are harnesses. Spec Kit, OpenSpec, AI-DLC, GSD, and Superpowers are workflows or methods.

Wrong: LangGraph vs AI-DLC
Better: LangGraph builds the agent app; AI-DLC governs delivery of that app.

What a harness does

An agent harness/runtime typically provides:

• model/provider connection;
• prompt and instruction loading;
• tool execution;
• file read/write;
• shell commands;
• memory;
• skills;
• subagents;
• approvals or safety controls;
• session/task lifecycle.

What a workflow framework does

A workflow framework typically defines:

• what artifact comes first;
• what counts as source of truth;
• when to ask questions;
• how to write a plan;
• how to split tasks;
• when to implement;
• how to review;
• what evidence proves done;
• how to archive or update docs.

What an agent app framework does

An agent app/orchestration framework typically provides:

• model abstraction;
• prompts and structured outputs;
• tool calling;
• retrievers and data integrations;
• state management;
• graph orchestration;
• checkpoints;
• human-in-the-loop;
• deployment/runtime hooks for AI apps.

Example stack

flowchart TB
    A[OpenAI API or local LLM] --> B[LangGraph agent app]
    B --> C[Hermes or Codex as coding harness]
    C --> D[OpenSpec workflow]
    D --> E[openspec/changes and openspec/specs]
    E --> F[Repo, tests, PR, CI]

The harness can change without changing the workflow. The workflow can change without changing the model.

Practical rule

Use this rule:

Choose a harness when you need execution capabilities. Choose a workflow when you need process discipline.

Need	Choose
Build AI app, RAG, or tool-calling agent	LangChain
Build long-running stateful agent system	LangGraph
Better CLI coding experience	Codex CLI or Claude Code
Open-source/customizable runtime	Hermes Agent
Strong spec-first feature workflow	Spec Kit
Lightweight change specs	OpenSpec
Enterprise approval/audit lifecycle	AWS AI-DLC
Multi-agent phase execution	GSD
TDD and review discipline	Superpowers