Do Claude Code agent teams keep working across a restart the way a crew does?

Tares keeps a crew working across a restart, because the plan, tickets, reviews and your answers stay on Tares, while Claude Code agent teams only partly survive one. Agent teams are experimental and off by default. Their task list persists, but /resume does not restore in-process teammates, so they suit work you finish in one sitting. With the software factory for Claude Code, a watch brings back stopped sessions and they carry on from the record.

How it works

The difference is where the state lives.

Claude Code agent teams

• A lead session creates a shared task list and assigns tasks to teammates.
• The task list files sit under ~/.claude/tasks and persist.
• The team config under ~/.claude/teams is removed when the session ends.
• /resume does not restore in-process teammates. After a restart you keep the task list, but the teammates are gone.
• Agent teams are experimental and off by default. They need no extra install once you turn them on.

A Tares crew

• You plan with Claude Code in /tares:spec. It interviews you and has a second model attack the plan. The plan is saved as a spec, milestones and tickets.
• The plan, tickets, reviews and your answers stay on Tares, local or Tares Cloud. The pull requests live on GitHub.
• A reviewer session checks every pull request to a fixed bar and posts its verdict. Builders merge only after that verdict and green CI.
• When sessions stop, a watch brings them back. They carry on from the record instead of starting over.

Because the record sits outside any one session, a restart of your machine or of Claude Code does not erase the plan or the decisions. For the full walkthrough, see Get started.

Set it up

To run a crew that carries on after a restart, follow these steps in order.

1. Check that Claude Code is signed in, that gh is signed in with write access to the repos the crew works in, and that git, jq and curl are installed.
2. Install the CLI with uv tool install tares or brew install glassflow/tap/tares.
3. On a local Tares, start Tares with tares up.
4. On a local Tares, open a second terminal and start the MCP endpoint every crew session shares. The crew reaches a local Tares only through this endpoint, so it comes before the crew.
5. On Tares Cloud, run tares login and pick your workspace in the browser. There is no MCP endpoint to start.
6. Give Tares access to GitHub: Settings > GitHub on a local Tares, or Connect GitHub on Tares Cloud.
7. Bring up the crew with tares factory crew up.
8. Plan a project with tares factory spec. Answer the orchestrator’s questions one at a time. Each comes with a recommendation.

gh auth login
uv tool install tares     # or: brew install glassflow/tap/tares
tares up
tares mcp --transport streamable-http
tares factory crew up
tares factory spec "what you want to build"

On Tares Cloud, replace tares up and the MCP line with tares login.

After a restart, the plan, tickets, reviews and your answers are still on Tares. The watch brings back the sessions that stopped, and they pick up from that record.

What you get

• A plan, tickets, reviews and your answers that outlive every session, kept on Tares and on GitHub.
• Stopped sessions that come back through the watch and carry on from the record, not from a blank start.
• An independent review of every pull request, with merges only after the reviewer’s verdict and green CI. You can require your yes before every merge.
• The crew runs on your machine on your Claude Code login, with no separate model key. The Codex check is optional.
• Cost stays tied to your Claude Code plan. Agent teams also use your Claude usage, and significantly more than a single session.

The side-by-side view of both approaches is on the software factory comparison page.

When something else fits

• Choose Claude Code agent teams when you want several sessions to debate a problem or split a feature while you watch, and you will finish in one sitting. They are built into Claude Code, though experimental and off by default.
• Choose subagents when the job fits inside one session, such as a research question or a focused review. They are the least setup, and they do not keep a plan beyond the session.
• The software factory comparison covers these and other options on the same seven questions.

Related