Choosing Your Harness Stack

The Decision

Build the stack from the bottom up:

1. Pick a base harness

Driving questions: where does the user live (terminal, IDE, browser), what's the team's primary model, how much customization will you do?

If...	Pick
Terminal-first, want hooks/sub-agents/plugins	Claude Code
Terminal-first, OpenAI ecosystem	Codex CLI
IDE-first, want IDE intimacy	Cursor
OSS, headless, agentic-OS-shaped	OpenHands
Web-first, lighter touch	Aider, Continue

2. Add an orchestration platform if needed

Driving question: do you need multi-agent / autopilot / federation?

If...	Add
Single user, single agent, interactive	Nothing — base harness is enough
Single user, want background workers / autopilot	ruflo
Multi-agent within one trust boundary	ruflo or build with sub-agents
Federated, cross-machine, zero-trust	ruflo (federation mode)
Long-horizon agentic OS	OpenHands or ruflo

3. Decide on a framework or SDK if you're building

Driving question: are you using a harness or building agent-shaped functionality into your own application?

If...	Use
Extending an existing harness	Plugins, sub-agents, slash commands, hooks
Building a Claude-powered app	Claude Agent SDK
Building a graph-shaped flow	LangGraph
Building a dialogue-shaped flow	AutoGen / AG2
Building a role-shaped flow	CrewAI
Building a Next.js / Node app	Mastra
Building on Gemini	Google ADK
Building on OpenAI	OpenAI Agents SDK

4. Standardize on MCP for tools

This is independent of the rest. Every modern harness supports MCP; tools you build for MCP work everywhere. Don't write tools in harness-specific formats unless you have a reason.

5. Keep configuration portable

Write CLAUDE (or equivalent) that's straight markdown — easy to convert to AGENTS, .cursorrules, etc. when the day comes to switch.

What Drives the Decision

Three factors dominate:

Team's existing tools and models. A team on OpenAI's platform with Azure deployment doesn't pick Claude Code first. A team with Anthropic credits and a customization mindset doesn't pick Codex CLI.
Customization appetite. Hook-rich harnesses pay off when you want to customize. They're overhead when you just want it to work.
Workload shape. Single-developer interactive coding → IDE/terminal harness. Long-horizon autonomous workflows → orchestration platform. Production agent service → continuous-loop platform.

A Worked Example

A small startup, 8 engineers, mixed Python/TypeScript codebase, on AWS with a Claude API contract.

Base harness: Claude Code (terminal, fits the team's vim/tmux workflow, hooks let them ship custom permission policies).
Orchestration: Adopt ruflo for the team's batch workflows (testgen, security audit). Don't use ruflo for interactive sessions — too much overhead.
Framework: For the customer-facing app's agent feature, use the Claude Agent SDK (matches the harness, lower context-switch cost).
MCP: All custom tools written as MCP servers from day one.
Memory: CLAUDE files committed to repos; team members can BYO .cursorrules if they prefer Cursor for some tasks.

A different team — a solo developer building agent features into a Next.js app — would pick differently: Cursor as harness (IDE-first), Mastra as framework (TypeScript-first), no orchestration platform (overhead).

Why This Decision Is Reversible (Mostly)

Switching costs are real but bounded:

Configuration files convert mechanically (CLAUDE → AGENTS is mostly find-and-replace).
MCP tools port without modification.
Sub-agent definitions convert with effort but mostly mechanically.
Trajectories and adapters don't port.

Investing in MCP and portable configuration is the practical way to lower switching costs. The ratchets are: trajectories, parametric adapters, deeply harness-specific hooks. These bind you to the harness you started with — appropriately so once you've committed.

Connections to Other Concepts

This concept synthesizes the course. Every concept you've read so far feeds the decision.

Direct connections:

the-2026-harness-landscape — The roster.
harness-vs-framework-vs-sdk — The category boundaries.
Claude Code vs. Codex CLI vs. Cursor — Harness comparison.
LangGraph vs. AutoGen vs. CrewAI — Framework comparison.
OpenAI Agents SDK, Mastra, and Google ADK — SDK comparison.
topology-selection-decision-tree — The orchestration-side decision.
memory-portability-across-harnesses — Why portable config matters.