Good morning. A capable coding model now fits on a laptop, agents are taking on day-long jobs, and today's build is the kind of agent I think every small business could use: one that chases overdue invoices and never sends a thing on its own.
BREAKING STORY
A good coding model now fits on your laptop
Thinking Machines released Inkling 2 Mini as open weights on 20 November, and the interesting part is where it runs. The whole model takes about 20GB, so any Mac with 32GB of memory can run it through Ollama, and it scored 79.2% on SWE-bench Verified in the company's own tests. If you would rather not host it, Artificial Analysis found its API, at $0.15 per million input tokens, about six times cheaper and four times faster per task than Qwen3.8 27B for the same score (it is not the cheapest on that index; GPT-6 Luna scores higher for less).
Code a client won't let leave your machine can now get a capable model's help without going anywhere. Pull it with Ollama this week, run it on three tasks from a live project, and compare the results with what you use now.
My take: I'd test this before paying for another hosted coding seat. A model that fits on a 32GB laptop and costs fifteen cents per million input tokens means a solo builder can keep client code on their own machine and give up very little quality.
MARKET TRENDS
Agents can now take on a day's work, if you can check it
METR's Q4 update, published on 20 November and covered by MIT Technology Review, puts current frontier models at a 50% time horizon of up to 27 hours on tasks with a verifiable answer. Read: they finish half of the jobs that take a skilled person about 27 hours. Across METR's tests that length doubles roughly every 3.6 months, which is fast enough to change what you can hand over within a year.
Judgement is another matter. On one test that asks for ratings of philosophical arguments, every public model still falls short of experts who spent 15 minutes on each, and AI judges overrated work from models as strong as or stronger than themselves.
How to make money from it: if I were selling client work, I'd sort a week of it into tasks with a test you can run and tasks that need taste. Hand the first pile to an agent, price the second as your own expertise, and sell faster turnaround on the checkable work.
TOOL REVIEW
InteraOne: one inbox for an AI agent and your team
What it does: InteraOne, which launched on Product Hunt on 20 November, puts an AI agent and your staff in one shared inbox for web chat, email, WhatsApp and Telegram. The agent answers first and hands the conversation to a person when it needs one, with tickets and a knowledge base alongside. It is open source and still in beta, and WhatsApp runs through Twilio, so you need a paid Twilio account and a Meta Business Portfolio.
Price: Starter is free with 50 AI messages, Pro is $9 a month for 500 messages and two human agents, and Pro Plus is $39 for 5,000 messages and ten human agents, as of 21 November.
Who it's for: shops, studios and service businesses whose customers message them on WhatsApp or Telegram.
Verdict: I'd give it a trial if you're a small shop answering customers on WhatsApp or Telegram; if you need a proven help desk, I'd still pick Chatwoot.
TODAY'S BUILD | AI AGENT
An AI agent that chases your overdue invoices
The context: Late payments reach every business, but they hurt most when you are small. A freelancer or a five-person studio pays rent, software and wages every month, so one client paying three weeks late can decide whether the month works. And the chasing is the job everyone puts off: keeping track of who owes what, remembering when to follow up, and writing the same awkward email again.
What it does: It keeps track of every unpaid invoice in your sheet and, each weekday at 8am, drafts a reminder in Gmail for any client who is overdue. It never sends anything on its own: you check the draft and press send.
Who it's for: Small businesses, freelancers and studios that invoice clients and would rather not chase them.
How it works: The agent follows a set rhythm. The first reminder comes three days after the due date, the next ones no more than once a week, and a final notice at thirty days. If a client is still unpaid after that, it stops chasing and flags the invoice to you, which is your cue to pick up the phone or talk to your accountant or lawyer.
Why I picked it: I think this is what a useful agent for a small team looks like: it does the tracking and the dull follow-ups, and hands the hard conversations back to you.
AI & WORK
AI saves more time when the people doing the work help redesign it
Nearly three in four workers now use AI agents at work: 74%, up from 64% last year, according to the Adecco Group's survey of 40,500 workers in 36 countries, published on 19 November. Use is no longer the question; value is. Workers who say AI is well built into their workflows report saving 20% more time a day than those who don't, and the sharpest split is about involvement. Of workers who helped redesign their work, 96% are willing to adapt, against 29% of those who weren't asked. Training is heading the wrong way, falling from 77% to 67% of workers this year, even as two-thirds say AI has changed the skills their job needs.
It is a survey of what workers say, run by a staffing group with a view on the answer, so read it as a signal. I think the useful lesson is the cheap one: before adding another AI tool, redesign one workflow with the people who do it, and tell them plainly what changes in their job.
OPEN SOURCE
A 145B-parameter model on one consumer GPU
Edge0, an open-source project with about 12,000 GitHub stars, released two mixture-of-experts models on 13 November, Edge0 145B and 120B, built to run on a single NVIDIA GPU. These models use only a few of their experts for each token, so Edge0 keeps the rest on your SSD and streams in the ones it predicts it will need next. On 20 November it added the same streaming for Apple silicon, so both now run on a 64GB Mac. The project reports 121 tokens a second on agent tasks with a 24GB RTX 5090 D, and coding scores above the Qwen model it was built from; those are its own numbers and have not been reproduced yet.
The framework is Apache-2.0, but the model licence needs written permission for paid model services and commercial coding assistants. I'd try it if you run agents on private code and want something bigger than the usual 30B local model.
WHAT GOT FUNDED YESTERDAY
Five startups that announced rounds on 20 November. The one I'd watch is Hoplite: coding agents that keep working after you close the laptop look like where developer tools are heading.
Hoplite · Ireland/US · Runs developers' AI coding agents in the cloud, so the work carries on after the laptop closes. Raised $5.5M, led by John and Patrick Collison, with Y Combinator.
Archer (formerly hackajob) · UK · An AI recruiting agent that finds and qualifies candidates for employers and suggests roles to job seekers. Raised $10M, led by Volition Capital.
Gotmoves · UK · Batches car deliveries into multi-stop driver routes, with live tracking, for dealers and fleets. Raised £2M, led by Midlands Engine Investment Fund II (Mercia Ventures).
SatSure · India · Turns satellite data into decisions for banks, insurers and governments, such as approving farm loans. Raised ₹60 crore (about $7M), the first close of its Series B, led by the Antariksh Venture Capital Fund.
NBS · Brazil · Lets people turn reais into digital dollars over Pix, and companies pay overseas suppliers in local currency. Raised R$11M (about $2M), led by BluStone.
TODAY'S GIGS
Five marketing and growth roles at AI companies that raised big this year. If I were applying, I'd start with Temporal's: it's remote, it went up this week, and the brief is demand generation aimed at startups, which is where agent infrastructure gets won.
Senior Startup Marketing Manager, Demand Generation @ Temporal (raised $550M, Sept) · Remote (listed under San Francisco) · Link
Senior Manager, Customer Marketing @ Fireworks AI (raised $1.5B, July) · Remote: United States · Link
Sr Field Marketing Manager @ Fireworks AI (raised $1.5B, July) · San Francisco or New York, hybrid · Link
Product Marketing Manager @ Arena (raised $200M, Oct) · SF Bay Area or New York, on-site · Link
Analyst Relations Manager @ Databricks (raised $5B, Aug) · United States · Link
QUICK HITS
Citi now makes customers liable for what a third-party AI agent does with their bank login, following JPMorgan and Bank of America (18 Nov); I'd read your bank's terms before giving an agent your login.
From 16 December, Etsy's terms let it and its partners train AI on seller listings, with no general opt-out (18 Nov).
Google opened a waitlist for Stitch Loop, an agent that watches your code and live metrics and opens pull requests for you to approve (18 Nov).
That's today. Hit reply and tell me what you're building. I read every email.
Dev

