Use one conversation per task

Why it helps

• Each conversation is a separate agent session. A new conversation starts the agent with a clean slate. Claude Code’s documentation explains that sessions are independent: “Each new session starts with a fresh context window.” Mixing tasks fills that window with details that don’t matter for the current one. Claude Code’s guide calls this “the kitchen sink session” and lists it as a common failure.
• Follow-ups belong together. OpenAI’s Codex guide recommends “one chat per coherent unit of work,” and adds that staying in the same chat for the same problem “preserves the reasoning trail.”
• Your list stays readable. The agent gives each new conversation a short label, so a focused conversation gets a name you can find later. Each one shows its own status, such as Working, and a Needs you note when the agent is waiting for you. You can mark it completed when it’s done.
• Settings stay separate. Each conversation keeps its own agent, model, effort, and permissions, and its own preview and running processes.

How to do it

1. Start a task with a new conversation. Select New conversation, the plus button at the top of the conversation list. If you’ve hidden the list, the same button sits beside the list button above the conversation. In the iPhone app, select the pencil button (New conversation) at the top of a project.
2. Keep follow-ups in it. Corrections, tweaks, and “one more thing” about the same task go in the same conversation.
3. Start another conversation when the topic changes. If your next message has nothing to do with the last reply, it’s a new task.
4. Mark it completed when you’re done. See Mark finished work as completed.

You can also ask an agent to start a conversation for a side task. Agents can start new conversations in your project, up to five for each request, and reply with a link to each one.

An example

One conversation:

• “Add a newsletter signup to the footer. Ask for an email address only.”
• “Make the button the same blue as the header.” (Same task, so it stays.)

A new conversation:

• “The site feels slow on my phone. Find out why and tell me what you’d change before changing anything.”

Common pitfalls

• One conversation for the whole project. Codex’s guide names “using one chat for an entire project instead of one chat per coherent outcome” as a common mistake.
• Splitting one task across many conversations. Each new agent has to rediscover what the last one learned. Stay put until the task is done.
• Two conversations changing the same page at the same time. They can overwrite each other’s work. See Run agents in parallel.
• Asking a quick unrelated question mid-task. Even a small detour stays in the conversation’s history. Ask it in a new conversation instead.