--- title: "Agentic tasks: multi-step work, Claude in Chrome, Excel…" description: "From answers to outcomes Until 2025, most people used Claude as a conversation: ask, read, copy. In 2026 Claude increasingly does multi-step work : it…" url: https://optimizeall.com/learn/mastering-claude/agentic-tasks-chrome-and-office updated: 2026-10-05 --- Mastering Claude (Anthropic) · Projects, Artifacts, personalisation and agentic tasks · lesson 10 of 20 · 16 min # Agentic tasks: multi-step work, Claude in Chrome, Excel, PowerPoint and Slack ## From answers to outcomes Until 2025, most people used Claude as a conversation: ask, read, copy. In 2026 Claude increasingly **does multi-step work**: it plans, uses your connected tools, creates files, checks its own output and comes back when it needs a decision. Anthropic first shipped this as **Claude Cowork** (generally available in April 2026, on web and mobile from July), then in September 2026 merged it into the main Claude conversation. You now simply describe the outcome, and Claude decides whether to answer, create a document, or work through a task, checking in with you and leaving you the final say. You can also ask for **recurring work**: "Every Monday, pull last week's numbers from the connected sheet and draft the team update." Claude starts the task on schedule without being asked. ## The surfaces where Claude now works | Surface | What it does | Good for | |---|---|---| | **Main conversation (agentic tasks)** | Multi-step work across connectors, files and code | Reports, research-to-deck, data clean-ups | | **Claude in Chrome** | A browser extension that lets Claude read and act on web pages you have open, with your permission | Filling repetitive web forms, comparing supplier pages, pulling data from dashboards without an API | | **Claude for Excel and PowerPoint (Microsoft 365)** | Works inside the workbook or deck, with cross-app support since March 2026 | Building models, explaining formulas, turning analysis into slides | | **Claude in Slack** | Delegate work from Slack (including tagging Claude in a thread) | Team requests, summaries, drafting replies | | **Computer use (research preview)** | Claude operates desktop apps via screenshots and clicks | Legacy tools with no connector | | **Mobile** | Start, steer and approve tasks from your phone | Checking in on long tasks | Availability differs by plan (many of these are paid-plan features) and by admin settings on Team and Enterprise. Check the help centre for your plan. ## Delegation that works: the task brief Agentic work fails for the same reasons delegation to people fails: vague goals, missing access, and no definition of done. Use a brief: ```text Outcome: A 1-page weekly performance update for the Noor Organics team. Inputs: Connected Google Sheet "Weekly KPIs"; last week's update in the Project. Steps you may take: read the sheet, calculate week-on-week changes, draft the update as a Claude Doc, create 1 chart. Do NOT: send, post or share anything; edit the sheet; contact anyone. Check-ins: ask me before any step not listed above. Definition of done: numbers match the sheet (show the cells used), 3 insights, 3 recommended actions, "Checks needed" list at the end. Schedule: every Monday 08:00 UK time, starting next week. ``` ## Guardrails for autonomy Agentic tools act with **your** permissions, so a mistake can be real. Anthropic builds protections (permission prompts, confirmations for consequential actions, prompt-injection defences, and, for plugins and skills, security scanning in beta), but you set the scope: - **Least privilege:** connect only the tools the task needs; prefer read-only access. - **Explicit stop rules:** "Do not send, delete, purchase, publish or submit." - **Approve consequential actions** yourself, especially in Claude in Chrome on sites where you are logged in (banking, ad accounts, admin consoles). Avoid those sites entirely for agentic browsing unless your organisation has approved it. - **Watch for prompt injection:** web pages and documents can contain instructions aimed at AI. Keep tasks narrow and review outputs. - **Log and review:** check what the task did before relying on it; for recurring tasks, review the first three runs closely. ## Worked example: a weekly report on autopilot, safely An e-commerce manager in Manchester connects Google Drive (read-only) and asks Claude to produce the Monday trading update from a KPI sheet. The first run shows the cells used; she finds Claude treated refunds as revenue because a column header was ambiguous, fixes the header, and adds a rule to the brief. From week three she spends ten minutes reviewing instead of ninety minutes compiling. Posting the update to Slack remains a human step. ## Hands-on 1. Choose a weekly task you currently do by hand that uses information Claude can reach. 2. Write a task brief with outcome, inputs, allowed steps, stop rules, check-ins and definition of done. 3. Run it once manually. Review every number and action. 4. Only then schedule it, and review the first three scheduled runs. 5. If you use Claude in Chrome, try it on a low-risk, logged-out task first (comparing public supplier pages). ## Pitfalls - Granting broad write access "to make it easier". - Scheduling a task before a successful supervised run. - Letting a browser agent loose on logged-in admin or payment pages. - Treating the output as final because the process was automated. ## How to measure success Time spent on the task falls sharply after the first few runs, error rates are no worse than the manual process (measure them), and no consequential action ever happens without an explicit human approval. ## Video lecture: Agentic tasks: multi-step work, Claude in Chrome, Excel, PowerPoint and Slack Lecture coming soon · 14 chapters · about 9 minutes. Read the full transcript below. 1. From answers to outcomes 2. Why this is a step change 3. How we got here 4. Where Claude works now 5. Delegation fails for human reasons 6. The task brief 7. Guardrails 8. Simple example 9. Worked example: Monday trading update 10. Pitfalls 11. Try this now 12. Watch me do it, part 1 13. Watch me do it, part 2 14. Recap and next step ## Lecture transcript ### From answers to outcomes For most of AI's short history, you asked and it answered. In twenty twenty six, Claude increasingly does the work, planning steps, using your tools, creating files and checking back with you. This lecture shows you where Claude can now act, how to delegate a task so it actually succeeds, and the guardrails that stop an eager assistant from making a very real mistake. ### Why this is a step change Why is this a step change? Because an answer saves you minutes, while a delegated task can save you hours. Compiling a weekly report, pulling numbers from three places, drafting the update, formatting the chart. That is the work agents now take on. But there is a catch. Once an assistant can act, mistakes become real. An email sent, a sheet overwritten, a form submitted. So think of it like handing a task to a capable new hire. The more you delegate, the more your briefing and your guardrails matter. ### How we got here A quick history. Anthropic launched Claude Cowork as a separate workspace for multi step tasks, generally available from April twenty twenty six and on web and mobile from July. Then in September it merged Cowork into the main Claude conversation. Today you simply describe the outcome you want. Claude decides whether to answer, create a document or work through a task, checks in with you along the way, and you keep the final say. You can even schedule recurring work, like every Monday, draft the team update from the KPI sheet. ### Where Claude works now Claude now works in several places. In the main conversation, it runs multi step tasks across connectors, files and code. Claude in Chrome is a browser extension that reads and acts on pages you have open, with your permission. Claude for Excel and PowerPoint works inside your workbook or deck, and since March can move work between them. In Slack you can delegate by tagging Claude in a thread. And computer use, still a research preview, lets Claude operate desktop apps through screenshots and clicks. Many of these are paid plan features, and admins control them on business plans. ### Delegation fails for human reasons Here is the key insight. Agentic tasks fail for the same reasons delegating to a person fails. The goal is vague. The assistant lacks access to what it needs. Or nobody defined what done looks like. So we use a task brief. ### The task brief A task brief has seven parts. The outcome, for example a one page weekly performance update. The inputs, such as a connected Google Sheet and last week's update. The steps Claude may take, like read the sheet, calculate changes, draft a Doc and create one chart. A do not list, do not send, post, share, edit the sheet or contact anyone. Check ins, ask me before any step not listed. A definition of done, numbers match the sheet with the cells shown, three insights, three actions, a checks needed list. And finally, a schedule, but only after a supervised run succeeds. ### Guardrails Agentic tools act with your permissions, so set the scope carefully. Connect only what the task needs, and prefer read only access. Write explicit stop rules. Approve consequential actions yourself, especially in Chrome on sites where you are logged in, like banking, ad accounts or admin consoles, which are best avoided entirely unless your organisation has approved it. Remember that web pages and documents can carry instructions aimed at AI. Anthropic builds defences, including permission prompts and security scanning for plugins, but narrow tasks and human review are your best protection. ### Simple example Let's start simple. Ask Claude, with read only access to one shared folder, to summarise every Friday what changed in that folder this week, as a draft Doc, and to send or share nothing. Run it once while you watch. Check the summary against the folder. Did it catch the renamed file? Did it misread anything? When the first run is right, schedule it. You now have a small, safe, useful agent, and a pattern you can extend to bigger jobs like the weekly trading update later in this lecture. ### Worked example: Monday trading update An e commerce manager in Manchester connects Google Drive with read only access and asks Claude to produce the Monday trading update from a KPI sheet. Because her brief required showing the cells used, she spots on the first run that refunds were counted as revenue, thanks to an ambiguous column header. She fixes the header, adds a rule to the brief, and reviews the next two runs closely. By week three, ninety minutes of compiling has become ten minutes of reviewing. Posting to Slack stays a human step. She also learned where not to use an agent. When she tried to have it reconcile supplier invoices in the finance system, the task needed judgement calls about disputed charges and access she did not want to grant. So that stayed manual, with Claude only drafting a summary of discrepancies from an exported file. Knowing where to stop is part of delegating well. ### Pitfalls Watch for four pitfalls. Granting broad write access to make it easier. Scheduling a task before it has succeeded under supervision. Letting a browser agent loose on logged in admin or payment pages. And treating an output as final because the process was automated. Automation changes who does the work. It does not change who is accountable. ### Try this now Try this now. Choose one weekly chore that uses information Claude can reach, a status update, a folder summary, a KPI digest. Write the seven part brief. Outcome, inputs, allowed steps, a do not list, check ins, a definition of done, and the schedule you want later. Run it once while you watch, and review every number and every action it proposes. Correct anything, and add a rule to the brief for each correction. Only after a clean supervised run should you schedule it, and even then, review the first three scheduled runs closely. ### Watch me do it, part 1 Let me set up the Monday trading update. First, connectors. I enable Google Drive with the narrowest access my admin allows, reading only. Then I paste the seven part brief. Outcome, a one page weekly update. Inputs, the Weekly KPIs sheet and last week's update. Allowed steps, read the sheet, calculate week on week changes, draft a Claude Doc, create one chart. Do not, send, post, share, edit the sheet or contact anyone. Check ins, ask before any unlisted step. Definition of done, numbers match the sheet with the cells shown, three insights, three actions, a checks needed list. I leave the schedule for later. Claude shows its plan, and I approve it. ### Watch me do it, part 2 The draft arrives with a cells used appendix, which is why I asked for it. I compare it with the sheet and spot the problem. Revenue includes column E, which is refunds, because its header just says adjustments. I rename the header in the sheet, and I add a rule to the brief, refunds are negative revenue, never add them. I rerun, and the numbers now match my own calculation. Only now do I set the schedule, every Monday at eight, UK time. For the next three Mondays I will review the output closely before I trust it. Posting to Slack stays with me. ### Recap and next step Recap. Claude can now carry out multi step work in chat, in Chrome, in Excel and PowerPoint, and in Slack. Brief it like a good manager, with outcome, inputs, allowed steps, stop rules, check ins and a definition of done. Keep access narrow and approve anything consequential. Your next step: pick one weekly task, write the brief, run it once under supervision, and only then put it on a schedule. ## Key takeaways - Since September 2026, Cowork is merged into Claude: describe the outcome and Claude can plan and carry out multi-step, even scheduled, work. - Claude also works in Chrome, Excel and PowerPoint, and Slack, with availability set by plan and admins. - Delegate with a task brief: outcome, inputs, allowed steps, do-not list, check-ins, definition of done, schedule. - Use least privilege, approve consequential actions, beware prompt injection and supervise runs before scheduling. ## Try it Write a task brief for one weekly task (outcome, inputs, allowed steps, do-not list, check-ins, definition of done). Run it once under supervision and record every correction you made before considering a schedule. - [Previous: Styles, custom instructions and memory](https://optimizeall.com/learn/mastering-claude/styles-instructions-and-memory) - [Next: Extended thinking: when deeper reasoning pays off](https://optimizeall.com/learn/mastering-claude/extended-thinking-when-and-how) - [All lessons of Mastering Claude (Anthropic)](https://optimizeall.com/learn/mastering-claude)