For most of the past year, an AI agent was something you watched: you gave it a task, it worked for a few minutes, you checked the result. Three launches in the last week point to a different model, where agents run for hours in the background. Notion rebuilt its agents so a single task can run for more than 24 hours. ChatGPT's Chrome extension can now read your open tabs and act in them. And OpenAI added 12-hour prompt caching so always-on agents do not pay full price every time they re-read their instructions.
Notion's agent harness before and after its October 2026 rebuild, in minutes.
Notion: agents rebuilt to run all day
Announced on 19 October, the sixth generation of Notion's agent harness powers both the personal Notion Agent and Custom Agents shared by a team.
- 24-hour tasks
- A single task can run for more than a day, up from five minutes.
- Own computer and browser
- Agents can browse, write scripts and run them to finish work.
- Longer memory
- Long conversations are compacted and large tool results offloaded, so agents keep their place.
- Review for me
- A separate model checks the agent's actions before it takes them.
- Auto router
- Picks a model per task; Notion says results match Claude Opus 5.5 and GPT-6 Astra at about 90% lower average cost on 20+ common tasks.
- More tools
- Notion MCP now connects agents to hundreds of tools, up from a few dozen in the summer.
- Bring your own AI
- Connect a ChatGPT Plus, Pro or Go account to your personal agent; bring-your-own-key is in alpha for Enterprise.
Pricing moves to credits. Business and Enterprise seats get 1,500 Standard credits a month, for work where Notion picks the model, and 500 Select credits, for work where you pick the model or run Custom Agents. The system is live for new customers and applies to existing ones from 2 November. An admin "AI Command Center" with per-member usage and spending limits is coming.
ChatGPT in Chrome: an agent in your tabs
OpenAI's Chrome extension, updated in mid-October, brings the ideas from its Atlas browser into the browser most people already use. Reports say Atlas itself has been shelved.
Chat
Read- Tag Chrome in a chat and ChatGPT can see your open tabs
- Compare products, summarise articles or pull figures from several pages at once
Work mode
Act- Fill in forms, send messages, download files and group tabs
- Works on sites where you are already signed in, such as Gmail or Salesforce
- Mistakes happen in your real accounts, so check before anything is submitted
Safeguards
Control- Asks before acting on each new site, by default
- Allow once, per site or everywhere; edit allow and block lists
- Works in Chrome, Edge, Brave, Opera and Vivaldi
OpenAI's 12-hour cache: cheaper always-on agents
Prompt caching lets an API reuse the start of a prompt, such as long instructions, tool definitions and reference documents, at a much lower "cached input" price. Until now, OpenAI's cache lasted around half an hour, so an agent that woke up every hour paid full price again. On 20 October OpenAI added an optional 12-hour cache in its API, aimed at persistent agents. Writing to the longer cache costs more than a normal request, so it pays off only if the agent really reuses the same context several times through the day.
No cache hitsCached all day
Hypothetical example: 50,000 tokens of instructions and tools re-read 24 times a day at $2 per million input tokens, with cached input at a tenth of the price. Check your model's actual rates.
Using long-running agents for data work
- 1
Give them a narrow, checkable job
"Refresh the weekly KPI page and flag any metric that moved more than 10%" works better than "keep an eye on the business".
- 2
Write the output somewhere reviewable
A Notion page, a pull request or a draft email, not a direct change to production data.
- 3
Cap the spend
Set credit or budget limits per agent before you leave it running overnight.
- 4
Keep a human approval step
For anything that sends, deletes or pays, require a click from a person.
- 5
Read the logs once a week
Long-running agents drift. A short review catches loops, wasted credits and odd behaviour early.
Check your understanding
1/3How long can a single Notion agent task now run?
Up from five minutes before the October rebuild.
2/3What does ChatGPT's Work mode in Chrome add to reading tabs?
By default it asks before acting on each new site.
3/3When does OpenAI's 12-hour cache save money?
Cache writes cost more, so the saving depends on real reuse.
You got 0 out of 3.
Related: when AI agents go off-script, the Personal Agent Protocol and AI in spreadsheets, October 2026.
Sources: Notion, "Agents, rebuilt to do more for less" (19 October 2026); Tbreak and Unite.AI on ChatGPT's Chrome extension (October 2026); OpenAI Developers' announcement of 12-hour prompt caching (20 October 2026) as reported by RuntimeWire. Notion's cost comparisons are its own.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me