Choosing a harness

Chat, Cowork, Claude Code, and Codex are different working environments. Choose based on the actions, files, and supervision your task needs.

Bottom line

Harness and model are separate choices. Choose the working environment first; choose extra model horsepower only when the task needs it.

01 // Three doors

Conversation, delegated work, or building

Talk

Best for questions, learning, brainstorming, drafting, and short back-and-forth work where you stay in the loop.

Delegate

Best for multi-step knowledge work that needs selected folders, files, connectors, skills, or scheduled tasks.

Build

Best for building software, editing repositories, running tests, and working inside technical project environments.

02 // Decision table

If you need...	Start with	Move up when...
An answer, explanation, or draft	Chat	The task needs files, tools, or several steps
A finished document or organized set of files	Cowork	The work becomes software or needs a developer loop
A website, app, script, or repository change	Code / Codex	The task stops being technical
A repeatable action on a schedule	Cowork	The automation needs a custom application or service

03 // Two dials

Controls the workspace: chat, files, tools, terminal, connectors, permissions, and how much the agent can do.

Controls the engine: capability, speed, usage, context, and specialized strengths.

A powerful model in the wrong harness still cannot access the files or actions the task needs.
A well-configured harness with good context can outperform a stronger model working blind.
Keep one source of truth for context so changing harnesses does not reset your work.

04 // Safety

01
Scope access
Give the harness only the folders, apps, and accounts needed for the current result.
02
Review the plan
Check what it intends to read, change, send, publish, or delete before letting a long task run.
03
Verify the output
Treat completion as a claim. Open the files, test the result, and inspect what changed.

Official sources