The three parts
| Part | Think of it as | Example |
|---|---|---|
| Agent | A reusable teammate with a role, instructions, skills, permissions, and budget | Release Writer |
| Runtime | The program or service that does the work | Codex CLI or Claude Code |
| Agent Run | One attempt, with its own progress, result, usage, and errors | Draft announcement, attempt 2 |
Example: a writer that handles every release
Create aRelease Writer Agent when launch copy is a recurring responsibility.
Give it the runtime and instructions it needs. Each announcement creates a new
Run, but the Agent keeps the same role and can reuse approved skills.
If the first draft is wrong, ask for another attempt. You still have one Agent,
but two separate Runs, so the result and cost of each attempt remain clear.
When to create another Agent
Create a new Agent when a repeated class of work needs a different role, runtime, instruction set, budget, or permission level. Use an existing Agent for a one-off request that already fits its job. A Run can start from Chat, an assigned Issue, a review request, an Automation, or a manual action. The starting point changes; the Agent/Run distinction does not.What a Run can show
- the answer or completion summary
- created files, reports, screenshots, or links
- usage and cost when the runtime reports them
- transcript, tool activity, logs, and final errors when available