William Geronco

Invoicefy 1.1 just got approved on the App Store it's a simple invoice maker for small service businesses. send an invoice or estimate from your phone in a minute, text the pay link to your client, and see who owes you at a glance free on iPhone and iPad apps.apple.com/us/app/invoic…

new ondeck and ocrolus survey of 860 small businesses. 61% use AI and 90% of those say it made a measurable difference. the ones not using it yet mostly are not against it. they just do not know which tool fits their business. that is the real gap

digitalocean shipped cuttlefish, an iphone app where you watch your agents run and approve or deny tool calls with the full command or diff in front of you. risky ones need face id. that is the right idea. approving should be easy to do and hard to do by accident

perplexity launched teammates this week. agents that keep a role, watch your tools, and bring finished work back for review. aravind says they already have about 10,000 running internally, 20x their headcount. the count is not the interesting part. someone still has to judge the work

box unveiled architect at boxworks. you describe a process in plain language and it designs the workflow on top of the files you already have, with a test pass and human review before anything goes live. building the agent is getting easy. proving it on real documents before production is still the hard part

snowflake shipped coco and cowork bots that keep working in the background. you set a high level goal like briefings before every customer meeting, and they only ping you when a decision or sign off is needed. that is the shape i want for most business agents. do the work. ask when it matters

merchantfloor just launched a product listing agent for small shopify stores. you drop in a photo and a few notes. it builds the images, description, seo, faqs, then waits for you to review before anything hits your admin. that last step is the whole product. agents that skip review are toys

most good automations start as something you got annoyed doing for the third time. write it down the second time and automate it the third

meta paid about 1.2 million in bug bounties on its muse agent and only 64 of 630 plus reports were valid. handing an agent your email, calendar and messages is a big deal. small teams should copy the habit not the budget. let people try to break it before your customers do

crowdstrike just shipped a kill switch for ai agents. that tells you where we are. agents are doing real work inside companies now and the question moved from can it do the job to how fast can you stop it when it goes sideways. if you're building agents build the off button first

intuit says as many as 60 percent of small businesses on quickbooks have invoices more than 30 days past due, and the ones using invoice automation get paid up to four days faster. chasing payments is the most boring job in a small business and it's the one that hurts cash flow the most. automate the reminders first

aws now lets you hand step functions workflows to an agent as a tool. the agent decides when to run it and what to pass in, and the workflow runs every step the way you defined it. this is how i want agents touching anything important. let the model pick the moment, don't let it improvise the process

google's new gemini 4 carbon is ranked number one on the artificial analysis index and the first people getting it are security teams, not developers. most of the jump is in finding, exploiting and patching vulnerabilities. a model that good at finding holes is good at using them too, so i get why they're starting with the defenders

google just opened its connected apps platform so any business can plug its app into search and gemini. people ask the ai to book or order something and it uses your app without them ever opening it. feels like early seo all over again. if customers start buying through an assistant you want to be one of the apps it can actually use

a harvard paper tracked 718 companies after they started using ai coding agents. lines of code went up 30% and pull requests 23%, but finished features barely moved. review time went up 49%. writing the code was never the slow part. if you add agents to your team, figure out who's going to review all of it first

OpenAI's new GPT-6.1 Terra is built for computer use. Filling out forms, clicking through desktop apps, the repetitive stuff. It's $1 per million input tokens and $5 output, half the price of the last version. That's the work most small businesses actually want off their plate, and price just stopped being the excuse

Researchers found a planted AGENTS.md file in someone's Downloads folder that told Codex to install a malicious npm package. In their tests Codex followed it over 90% of the time. No code at all, just a text file the agent trusted. Give your coding agent one project folder, not your Downloads or your whole home directory

Nous raised $17M for an agent that handles household admin. It moves people to cheaper insurance and broadband, cancels subscriptions and books tradespeople, and almost 250k UK homes use it. The big thing is it acts instead of suggesting. That's a trust problem more than a tech one

BOB out of Miami just raised $5.5M to be an AI finance employee for small businesses. You text it on WhatsApp or SMS and it keeps the books, sends invoices and takes payments. No dashboard to learn. The roofer doesn't want another app, they want the work done