A stack map for agent work

How to split work between Cursor, Claude Code, Codex, subagents, and the human who owns product decisions.

What you get

A short decision map that connects every piece of work to the right tool

Who it is for

For anyone who collects smart tools but still opens every task in the wrong one.

1. The rule

The question is not which tool is best. The question is which piece of work you are handing off right now.

A quick edit is not task management.
Deep research is not a product decision.
A recurring process is not another one-off prompt.
A good tool starts and ends at a clear point.

2. Division of responsibility

Quick edit in a file: Cursor.
Multi-file task with tests: Claude Code.
A process that repeats and you want to keep: Codex or a Skill.
Research, review, or a clean room: a subagent.
Product, money, customers, and permissions: a human.

3. Questions to ask

Do you need a quick edit or a managed run?
Do you need project memory or a clean room?
Is there an external action that needs approval?
Does the action repeat enough to become a skill?
Who brings the proof at the end?

4. Map template

Copy box

Task: [what needs to happen]

Editing:
- Tool: [Cursor / Claude Code]

Running and tests:
- Tool: [Claude Code]

Research / review:
- Tool: [Subagent]

Recurring process:
- Tool: [Codex / Skill]

Completion evidence:
- [test / screenshot / log / link]

Decisions you do not hand off:
- [product / customer / money / permissions]

5. Signs the stack is working

Every tool knows where it starts and where it stops.
No single tool holds every permission.
The human stays on business decisions.
Tests and evidence close tasks.
No tool switch happens without a reason.

How to use this now

Write this map once. After that, every task starts from a division of responsibility, not from hesitating over another logo.

Explore

Tools