What is an agent?
Think of an agent as a tool that does more than generate text, it can perform tasks like drafting replies, summaries or running scheduled work for you.
← The Human Bit Weekly
11 October 2026 · 6 min read
Let an AI draft a summary or reply, but you decide when and whether it leaves your hands.
The 60-second version
It’s easier to hit send fast. It’s still your job to pause where it counts.
- What changed
- Try itMCP grew up: stateless, versioned and wired for enterprise identity.
- Test carefullyClaude Cowork now runs on web and mobile. The agent moved off your desk, and off your machine.
- Worth watchingClaude went to work for government, with the audit trail public services need.
- What to do
- Make sure the draft contains only what you’d be willing to say or sign your name to, with nothing sensitive you missed in the review.
- What stays yours
- You decide if the AI’s draft is ready to send, and you’re accountable for what goes out.
Rather listen?
0:006:00
New to this?First time with AI?
What is stateless?
Stateless means the AI treats each request as separate, so it doesn’t remember details from run to run. Good for security and privacy, but saves less automatically.
What is an audit trail?
An audit trail is a record of what happened and who approved or changed each step. It matters when someone will need proof you really reviewed and sent something yourself.
Start here
Three things that changed
01Try itNewly announced
MCP grew up: stateless, versioned and wired for enterprise identity.
Stateless protocols and enterprise-ready permission checks mean automated reply or summary tools can finally reflect who should see what. This is only a gain if you wire it up; most people still rely on simple copy-paste and basic sharing.
See the detail and the sources →02Test carefullyNewly announced
Claude Cowork now runs on web and mobile. The agent moved off your desk, and off your machine.
An agent that drafts for you in the cloud means it can fill in while you’re away. But as your work moves into browser and mobile sessions, you give up pinning context and storage to your own machine, so knowing what lands in cloud storage before sending matters more than ever.
See the detail and the sources →03Worth watchingNewly announced
Claude went to work for government, with the audit trail public services need.
The government-focused release is a live test of what happens when draft approvals and audits aren't optional. Many private teams still work on a ‘whoever replies first’ basis, but a record of approvals is no longer just bureaucracy for public sector teams.
See the detail and the sources →
The one that mattersIt’s easier to hit send fast. It’s still your job to pause where it counts.
AI can now prepare a draft even when you’re not at your desk and lets you call up work wherever you are. The new plumbing removes tedious steps, but it also blurs who is actually saying what and when, especially as sessions and files live in the cloud, or approval becomes a technical formality. This is the week to try AI as your draft-preparer or summariser, then set a rule: nothing leaves your hands or gets a green light unless you say so, for a reason you can name.
The full take: what is new, the makers' claim, where it fits, what is still unknown
What's genuinely new
Cloud-based, identity-aware agents draft or summarise for you, on any device. Audit-ready setups move accountability upstream to the moment you approve.
What the makers claim
The idea that wiring up identity or using cloud agents is automatically an upgrade for every user, and that audit trails are only a concern for public sector.
Where it helps
Preparing a response to a tricky email, or a summary of meeting notes, on a day when you’re not at your usual machine.
Where it doesn't
Any workflow where you must act immediately and can’t review, like live support chat, or where all your files must stay on your own device.
Anthropic brought the MCP 2026-07-28 revision to Claude. The protocol's core became stateless request-response, which suits serverless and edge deployment; MCP Apps and Tasks moved under a versioned extensions framework for interactive UI and long-running work; and authorisation now follows production OAuth 2.0 and OIDC, connecting to enterprise identity systems. Claude's connector directory already lists more than 950 MCP servers.
Make sure the draft contains only what you’d be willing to say or sign your name to, with nothing sensitive you missed in the review.
- If cloud-based, agent-prepared drafts introduce new blind spots about where work happens or who is responsible.
- Whether people actually spot errors or issues in AI-prepared drafts before sending, once the workflow feels routine.
Try this
One thing to try this week
Draft a tricky reply or summary with an agent, but only you send it.
AI prepares a draft reply to a client email or a summary of project notes while you’re away from your main device.
You decide if the draft fits your intent and if it’s safe or complete enough to send. Only you, not the AI, hit send or save.
- 1
Find a recent client email or meeting transcript that needs a careful reply or summary.
- 2
Open your agent or AI tool on web or mobile and ask it to draft a reply or summary. Save the draft in the session, don’t send automatically.
- 3
Read the draft yourself. Highlight changes that make it clearer, safer, or more accurate before you copy, paste or send.
Human checkpoint
Make sure the draft contains only what you’d be willing to say or sign your name to, with nothing sensitive you missed in the review.
The Human Lens · what stays yours
Knowing what counts as a real approval stays your job for now.
AI does the typing, but it can’t know what’s on the line in a tricky client reply or a project summary you sign off. Only you know the shortcut you can take and the one you shouldn’t. Today’s workflow is about knowing when a draft is close enough, and when a bad send is still on you.
One question to carry into your week
Before you reply to that tricky client or send that summary, what still needs your yes or no before it leaves your hands?
See you next Monday.