# Continue the thread or start a new one?

Source: https://ryansam.com/agent-auto-continue/continue-or-new-thread/
Published: 2026-10-02 Updated: 2026-10-02

Your agent stopped.
A usage limit, a long lunch, or it just finished a step.
Do you type Continue, or start fresh?

My rule: same task, same thread.
New task, new thread, the moment it's new.
Anthropic's rule of thumb is the same: [start a new task, start a new session](https://claude.com/blog/using-claude-code-session-management-and-1m-context).

## What a long thread costs

Everything in the thread is what the model sees: your messages, every file it read, every tool output.
Anthropic calls the cost context rot: [performance degrades as context grows](https://claude.com/blog/using-claude-code-session-management-and-1m-context), and older, irrelevant content starts to distract.
A fresh thread costs too: for related work it rereads the files you just changed, which is slower and pricier (same post).

## Same thread or new thread

| Same thread if... | New thread if... |
| --- | --- |
| The next message needs the earlier ones ("do the rest the same way") | It'd make sense to a stranger ("add dark mode to settings") |
| It stopped mid-task | The task's done |
| It remembers the plan and what you decided | It forgets what you settled, or retries a failed fix |
| It's just long, and `/compact` would tidy it | It compacted and lost something you need |

One grey area: docs for the feature you just built.
Anthropic's post suggests a subagent for that; I stay put while the thread's still sharp.

## 23:40, three test files to go

It's 23:40 and Claude Code hits its session limit halfway through fixing your flaky auth tests.
Four test files done, three to go.

- **Continue.** "Do the other three the same way" means nothing without the first four.
  If you'd rather not wait up, Claude Code's [built-in auto-continue](https://ryansam.com/agent-auto-continue/claude-code/) picks it up in the same session after the reset (leave it open).
- **Then a new thread.** Next morning the migration's merged and you want dark mode on the settings page.
  The auth test history only gets in the way.

## Starting fresh without losing anything

- **Claude Code.** [`/clear`](https://code.claude.com/docs/en/commands) starts a new conversation with empty context (`/new` is an alias).
  The old one stays: `/resume` brings it back, and [`claude --continue`](https://code.claude.com/docs/en/cli-reference) reopens the most recent one in the directory.
  `/compact` summarises but keeps the same conversation.
- **Codex CLI.** [`/new`](https://learn.chatgpt.com/docs/developer-commands?surface=cli#start-a-new-chat-with-new) starts a fresh chat in the same CLI session, [`/compact`](https://learn.chatgpt.com/docs/developer-commands?surface=cli#keep-transcripts-lean-with-compact) swaps earlier turns for a summary, and [`codex resume`](https://learn.chatgpt.com/docs/developer-commands?surface=cli#codex-resume) reopens an earlier chat.

Claude Code also [compacts by itself](https://code.claude.com/docs/en/how-claude-code-works#when-context-fills-up) as context fills up, and early instructions may not survive.
Anthropic's [example](https://claude.com/blog/using-claude-code-session-management-and-1m-context): after long debugging, "now fix that other warning we saw in bar.ts" lands on a summary that may have dropped it.

Before switching, ask the old thread for a handoff note and make it the new thread's first message:

```text
Write a handoff note for a fresh session that will pick this up.
Under 200 words: the goal, what's done, what's left,
the files that matter, what we decided
and which approaches we ruled out.
Skip what the code already says.
```

At a Claude Code usage limit the old thread can't answer, so write it yourself under the same headings.
Anything every thread should know belongs in `CLAUDE.md`, [not in the chat](https://academy.claude.com/courses/claude-code-101/context-management).

## Where the app fits

[Agent Auto-Continue](https://ryansam.com/agent-auto-continue/) is a free macOS menu-bar app that sends a message into an agent's conversation at a time you pick, or once the limit lifts.
It only sends into a conversation that already exists, and never starts one.

When a fresh thread's the right call, start it yourself with the handoff note, then pick it in the app.
For what to send, see [what to say instead of Continue](https://ryansam.com/agent-auto-continue/what-to-say/).

![The Agent Auto-Continue conversation picker, Choose a thread, listing eight conversations newest first from Claude Desktop, Codex, Claude Code, OpenCode, T3 Code and ChatGPT (Codex), each with its agent mark, project and age, from Draft the 2.4 release notes 8 minutes ago to Add rate limit headers a day ago.](https://ryansam.com/_astro/picker-light.BDGHi_uQ.png)

*Every agent's conversations in one list, newest first, so you pick the one you already have.*

## What it can't decide for you

- **Whether the thread's gone soft.** That's your call.
- **Whether the task's done.** No agent reports that, so the app never guesses.
  A turn limit, a stop phrase or Stop ends a chain.
- **That you've moved on.** Before each send it re-checks that conversation and won't send if it's gone, archived or you've replied in it since.
  A reply in your new thread isn't a reply in the old one, so cancel or stop the old schedule from its row.