AI Compass / Knowledge cards

Long runs and parallel work

Claude Code's built-in tools

Checked 17 October 2026 · Michael Engel · Automation · Costs

Some tasks are too broad for a single chat, such as checking every file in a project for the same problem. According to the docs, a workflow is a script that Claude writes and that controls many subagents (helper agents) at the same time. Ultracode is a setting that lets Claude decide for itself when to plan a workflow. Three more tools help with long runs: auto mode, in which a second model checks every action instead of you; worktrees, separate working copies of the same repo; and the monitor, which watches logs in the background.

What it is for

Large tasks that split up well run in parallel and in the background. You don’t have to approve every action one by one.

Steps

1. Check whether the task fits: it splits into independent parts, and a cross-check is worth the effort. Not suitable are step-by-step work, reworking a single file, debugging, and tasks under five minutes.
2. Clarify the goal, constraints and desired output beforehand, and let Claude ask follow-up questions.
3. Trigger a workflow: write the word ultracode in the prompt or explicitly ask for a workflow. /effort ultracode turns it on for the whole session (according to the docs). To get to know the feature, there is the ready-made workflow /deep-research.
4. Have facts checked adversarially: several skeptics try to refute a claim, and the majority decides. For decisions, form a panel: one advocate and one skeptic per option, several judges with their own perspective, such as cost, time and brand, and a chief judge.
5. Watch the cost: /workflows shows progress and tokens per agent, and you can pause (p) or stop (x) a run there. Test on a small sample first. Afterwards turn ultracode off again with /effort ultracode off.
6. Run long jobs in auto mode instead of --dangerously-skip-permissions: a classifier checks every action and blocks risky ones, such as mass deletion or data leaks. Have variants built in their own worktrees (claude --worktree <name>, or “use worktrees for your agents” for subagents), then adopt the best one.
7. Use the monitor: first let Claude think about which anomalies are possible, then have it watch logs in the background and report only when something stands out. Save a successful workflow: the s key in /workflows turns it into your own command.

Limits

• Cost: according to the docs, every run counts against your quota. With ultracode you reach session and weekly limits faster. AI LABS cites about 1 million tokens for a small research task. That is the creator’s figure.
• The way it works has changed since June. In the videos it was 10 parallel agents and a keyword. According to the docs, up to 16 can now run at once (configurable, fewer with fewer CPU cores) and at most 1,000 per run. By default Claude plans considerably fewer agents. Ultracode is a setting alongside the effort level. The keyword only works in an open session, not with claude -p or scheduled tasks. On the Pro plan, workflows must first be switched on in /config.
• Auto mode is not a guarantee, and the classifier can be wrong. According to the docs, it has been the startup mode in the terminal since version 2.1.283. Run agents with write access in test folders first.
• Agent teams, meaning several sessions that coordinate with each other, are experimental and off by default according to the docs (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1). Try them only once the steps above work for you.
• A local /loop only runs while the session is open. According to the docs, recurring tasks expire after 7 days. For continuous operation you need the cloud or a scheduler.
• The model details in the June videos (Opus 4.8 as the strongest model) are out of date.

Links

• Let Claude work until the goal is met – with /goal: the tool for specifying an end state. The video of 29 June names it as the first step.
• Separate building and checking: the review commands for the result of a long run.
• Claude Code in the cloud: where runs keep going when your computer is off.
• Model choice: cost per task: smaller models for stages that don’t need the strongest one.

Sources

• Jack Roberts: Claude just dropped UltraCode… Unlock SuperPowers, 9 June 2026, 2:44–6:00, 10:49–11:22, 12:59, 17:15–18:21. Sponsored by Vanta, plus self-promotion
• AI LABS: These Claude Workflows Turn Claude Code Into A Whole OS, 3 June 2026, 4:15–6:57, 8:35. Self-promotion for the paid community
• AI LABS: The Top Claude Code Features We Use, 29 June 2026, 2:42–11:55. Sponsored by ScaleKit; the sponsor runs through the first minutes as part of the story (0:32–2:42)
• Claude Code docs: Workflows, Models and effort (ultracode), Permission modes (auto mode), Worktrees, Tools (monitor), Agent teams, Scheduled tasks (/loop), retrieved 17 Oct 2026

Status

Checked 17 Oct 2026 · Maturity new (built-in tools available and described in the docs; agent teams experimental; cost figures from creators only)

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab