The Helix Loop / Helix Loop

Agent Watch #3: Hand over the work, keep the final call

Agent Watch #3: What changed, 19 to 25 October 2026

The Helix Loop, 27 October 2026

See all 11 slides as a PDF.

What's important

Closing your laptop no longer has to stop the work. Claude Enterprise now combines chat and Cowork in one conversation across web, desktop and mobile. Tasks run in isolated cloud sessions and continue after you close your laptop. My judgement: use this in Augment mode for a report, analysis or deck with a clear review point. Decide what a useful result looks like before handing it over. Official source.

What stays human: you set the task, review the output and decide what is ready to use. Work continuing in the background still needs someone responsible for the result.

What changed for knowledge workers this week

For 19–25 October, these are the changes I would connect to everyday work.

  • Hand over a longer task. Claude Enterprise brings chat and Cowork together, with cloud tasks continuing across devices. Admins can control model access by role and spending at organisation, group and member levels. People set the brief and budgets, then review the work. Official source.
  • Prepare work before asking for approval. Claude Duet opens a research preview by waitlist for US Pro and Max subscribers. It monitors connected apps and prepares work in the background, asking before sending messages, paying, signing or deleting. People grant access and approve those actions. Official source.
  • Check an agent's proposed actions. Notion Agent and Custom Agents gain rebuilt infrastructure, their own computer and browser, and improved context management for longer tasks. Review for me uses a separate model to check actions before execution. People still set access and decide what to trust. Official source.
  • Ask questions across the data. Gemini in Connected Sheets can query the full BigQuery dataset, answer follow-up questions and build refreshable charts, pivots and dashboards. Rollout starts for eligible plans and users with query permission. You check definitions, interpretation and the conclusion. Official source.
  • Keep a record of AI work. Google Vault can filter and export Gemini in Workspace conversations by user or organisational unit and date range. Exports include prompts, responses, generated attachments and referenced-file links. People decide the scope and how to handle the material. Official source.
  • Distinguish drafting from sending. Grok Bot can allow explicitly requested email or Slack sends immediately. Routine sends still wait as drafts, and draft-only requests no longer pass through as sends. Write the instruction clearly; people own permissions and sensitive actions. Official source.

For coders, briefly

  • Claude Code adds Quick, Standard or Deep Code Review levels; developers choose the review and approve changes. Official source.
  • GitHub's app in Teams can investigate issues and create pull requests from notifications; people review the proposed fix. Official source.

What coders are doing that you should borrow

Give the draft fresh eyes — Augment

Coders on X are discussing review loops where another agent checks work with fresh context.

For your work: a client brief needs a reader who can spot what the writer assumed. Try: open a fresh chat and ask it to read your draft as a client, board member or regulator. Ask for unsupported claims and missing context. Augment means AI carries several steps; I review. You assess the objections and make the high-stakes calls.

Set the stopping points — Adapt

Coders are adopting background jobs that pause to ask for input or approval.

For your work: research or reconciliation needs a clear boundary before an action affects someone else. Try: specify what the agent may prepare and where it must stop: send, pay or sign. Adapt means AI works within rules I set; I stay on call. You set permissions and answer when judgement is needed.

Keep your method in files you own — Assist

Coders are sharing ways to carry skills, rules and connection settings between tools.

For your work: keep tone, source requirements and sign-off rules together in a plain file. Try: write the method for a recurring campaign update and ask AI to find gaps. Assist means AI helps me think; I decide. You maintain the method and choose which suggestions to accept.

One thing to try this week

Take a draft brief and its approved sources. Give them to a fresh chat. Ask: “What would the intended reader challenge? Which claims lack support? What context is missing?” Compare the critique with the sources yourself. Change the draft only where you agree. Keep the useful review instructions for next time.

Follow along

Explore the Agent Watch tracker and What changed ledger.

Subscribe to The Helix Loop and get Agent Watch every week.

If you want to practise these modes on your own work, Craft is where I work through them.

Suhit