An AI agent that can click "buy" for you is now open to anyone with a Northwind subscription. Here's what it can do, where it stops, and how to keep the final say.
The Frequency
Northwind's new agent shops, books and fills in forms for you
What happened. Northwind released Delegate on Monday, an AI agent that works inside its browser and finishes multi-step tasks for you: comparing prices, booking appointments, filling in forms. It's included with Northwind Plus ($20 a month) in the US and UK, with a waitlist everywhere else.
How it works.
- You describe the task in plain words: "book a haircut near me on Saturday afternoon".
- Delegate opens pages, reads them and clicks through like a person would, while you watch in a side panel.
- Before anything that spends money, sends a message or shares personal details, it stops and asks you to approve.
- It can't see your saved passwords. When a site asks you to sign in, you do it yourself.
What it means for you. This is the first big consumer app that lets AI finish a purchase, not just suggest one. It will save time on chores, but agents still misread pages: Northwind's own launch notes say Delegate completes about 7 in 10 test tasks without help (company-reported). The other 3 need you.
What to do. If you try it, start with low-stakes tasks, a restaurant booking rather than a flight, and keep the approval step on. Check your card statement in the first week.
SOURCE Northwind blog
The Scan
Lumen Labs cut the price of its mid-size model by 60%. Developers now pay $0.40 per million input tokens for Lumen Mid, down from $1. Expect AI features in the apps you use to get cheaper, or more generous on free plans.
SOURCE Lumen Labs pricing
Quillstack reportedly raised $40M for AI bookkeeping. The Ledger reports the round; Quillstack hasn't confirmed it. If it's true, AI for accountants is moving from experiments to real budgets.
SOURCE The Ledger
Kestrel-7B is the top trending model on Hugging Face. The open-weight model runs on a recent laptop, and its creators' own tests put it level with models twice its size at coding. Free models you can run privately keep closing the gap with paid ones.
SOURCE Kestrel model card
Lumen Chat was down for about two hours on Monday. Lumen's status page blamed a database fault and said no user data was lost or exposed. If your team relies on one AI tool, keep a second one ready.
SOURCE Lumen status page
Preset
The meeting-notes prompt
Turns a messy meeting transcript into notes people actually read. Paste the transcript into any AI assistant, then add:
Summarize this meeting in three parts: decisions made, open questions, and action items with an owner and a due date. Use only what was said. If an owner or a date wasn't stated, write "not set" instead of guessing.
The catch: don't paste confidential meetings into a tool your company hasn't approved.
Tested by Tuned In
Decoded
Small models that check their own work
The question: Can a small AI model catch its own mistakes without help from a bigger one?
What they did: Researchers trained a 3-billion-parameter model to write an answer, re-read it, and flag the steps it wasn't sure about.
What they found: On grade-school math problems, the self-checking version got 14% more answers right than the same model without the check.
The catch: It's a preprint, it was only tested on math, and checking makes answers about twice as slow.
Why you should care: Cheap models that know when they're unsure could run on your phone and say "I don't know" instead of guessing.
SOURCE Preprint on arXiv
The Reading
62%
The share of developers in the Stackwise 2026 survey who say they ship AI-written code to production at least once a week, up from 41% a year earlier. The question is no longer whether AI writes code, but who reviews it.
SOURCE Stackwise survey
Off Air
Asked by a user to write a haiku about Monday's outage, Lumen Chat declined, saying it needed "more uptime for poetry". Lumen's official account shared the screenshot itself.
SOURCE Lumen on X