We're launching Hollr 🇨🇦 When something breaks at home, you shouldn't wait days just to learn what's wrong. Hollr puts a licensed technican on a live video call with you in minutes. $34 for 15 min, only if they connect. Early access 👇 gethollr.ca/x

Founder things nobody puts in the @ycombinator videos: 1. You rewrite the same onboarding email five times and send version two 2. The best idea of the day shows up at 10:30 PM, never 10:30 AM 3. "Quick call" means 45 minutes 4. The customer who complains the most is often the one who refers you 5. You check the dashboard more than your texts Number 2 is happening right now. Which one is you this week?

Every client agent project at @Arosplatforms goes through the same 5 stages: 1. "Can it do everything?" 2. "Okay, can it do this one thing?" 3. "Why did it do that one thing?" 4. "Can it do it more like Linda in accounting does it?" 5. "Great, now can it do everything?" Stage 4 is where the real spec finally shows up. Same arc whether it runs on @OpenAI or @AnthropicAI models. Which stage are your clients in right now?

Rule at @Arosplatforms that took us too long to land on: The agent that did the work never grades it. A second agent gets the original ask and the output, and answers PASS or FAIL, plus one line on why. Notes are allowed, but each one needs a name next to it, and if it's still open at the next review it turns into a FAIL. @alexcovo_eth moved his reviewer to plain PASS or FAIL tonight too. Do your agents check their own work, or does something else check it?

4 checks before I leave a coding agent running overnight: 1. It works on its own branch, never main 2. It has a step limit, not just a time limit 3. Tests run on every commit, so I can see exactly where it went sideways 4. No access to anything with a password in it Waking up to 40 commits is fun. Waking up to 40 commits on main is a different morning. Same rules whether it's @claudeai Claude Code or @OpenAIDevs Codex. What would you add before bed?

7 things in the current MCP spec to know before you build a server, the protocol @AnthropicAI open sourced: 1. Tools: functions the model can execute 2. Resources: context and data for the user or the model 3. Prompts: templated messages and workflows for users 4. Elicitation: the server can ask the user for more info mid task 5. Extensions are opt in, both sides agree at initialization 6. Tasks extension: long running jobs with polling and durable handles 7. Tool annotations are untrusted unless the server is trusted Which of these does your server actually use today?

How we handle model updates on client agents at @Arosplatforms, with @OpenAI shipping again today: We never switch the same day. We rerun a batch of last week's real tasks on the new model, put old and new outputs side by side, and only switch if the boring ones still come out right. New models are usually better at the hard stuff. The surprises show up in the easy stuff. How long do you wait before switching a model in production?

Legacy code is where this @OpenAI update gets interesting for me. The scary systems at banks and governments aren't the new ones, they're the COBOL nobody left in the building can read. If I were testing it at @Arosplatforms, I wouldn't start with a migration. I'd hand it one old batch job, ask it to write down what the job actually does, then check that against the one person who still remembers. Reading first, rewriting later. Who here still has a Fortran or VB app quietly running something important?

How @Arosplatforms hands API keys to the agents we build for clients, with stolen key stories going around this week: 1. One key per agent, so a leak shows up as one line on the bill 2. A hard credit limit on every key (@OpenRouter lets you set one per key) 3. Keys live on the machine that runs the agent, never in the repo or the prompt 4. A daily spend alert at half the usual amount 5. A new key whenever someone with access leaves What's the one you'd skip because it feels like overkill?

Small thing we added to every client agent at @Arosplatforms: At the end of each day it writes the business owner a 5 line note in plain English. What it did, what it skipped, and what it wasn't sure about. Turns out owners would rather read a short note than open a dashboard. And the "wasn't sure about" line is where most of our fixes come from. Same setup whether the agent runs on @AnthropicAI or @OpenAI models. What would you want your agent to tell you every day?

6 bundled Claude Code skills worth trying this week, straight from the official @claudeai docs: 1. /doctor runs a setup checkup (duplicate installs, slow hooks, unused skills and MCP servers vs their context cost) and asks before changing anything 2. /doctor prompt-audit flags outdated or conflicting instructions in your CLAUDE.md files and skills 3. /verify builds and runs your app to confirm a change works, without falling back to tests or type checks 4. /run-skill-generator records how your project builds and launches, so /run and /verify follow that recipe every time 5. /claude-api cost-optimize profiles where your Claude API spend goes and suggests savings one change at a time 6. /skill-doctor shows what each skill costs in context and how often it gets used Which one would you run first?

Something we changed on client agents at @Arosplatforms this month: Every agent gets a short "never do this" list in plain English, right next to its instructions. Not the tool permissions. Those already block things. This is for the stuff that's technically possible but wrong, like replying to an angry customer or promising a delivery date. The list keeps getting longer, which is how I know it's working. @OpenAI What's the first thing on your agent's never list?

A rule we set early at @gethollrr: pros advise over video, but live electrical panels and gas lines always get an in-person referral, never a walkthrough. Every licence is checked with the issuing body in its province before a pro takes a call. What would make you trust a stranger on video with your home?

A tiny fix that cleared up a lot of weird agent mistakes for us at @Arosplatforms: Every client agent now gets today's date, the client's time zone and their business hours in the first lines of its instructions. Without it, "follow up Friday" means whichever Friday the model assumes, and a "morning" reminder can go out at 3 AM local time. Every model we use needed it, @GoogleDeepMind ones included. What's the most basic context your agent was missing?

5 questions I ask before we automate anything for a client at @Arosplatforms: 1. Who does this today, and how many hours a week does it take? 2. What happens when it gets it wrong? 3. Who notices first, the client or their customer? 4. Can it be undone in one step? 5. Does a person actually read the output? If the answer to #2 is "someone gets paid the wrong amount," it starts as a draft a person approves, not an action. And sometimes the honest answer is a plain @zapier zap, not an agent. What question would you add?

One button we keep switched off on client machines at @Arosplatforms: "always allow" in coding agents. Approving a whole command family once feels like saving time. But a line of text planted in a repo or a ticket can steer the agent into asking for exactly that rule, and a tired person clicks yes. So it's named commands only, no wildcards, and the allow list gets a look every Friday. Same rule in Claude Code, Codex and @cursor_ai. Which command do you always end up allowing?

Things agents say with total confidence: 1. "Task complete" (it was not) 2. "I couldn't find the file" (it was in Downloads) 3. "I've updated the spreadsheet" (a new one, called Copy of) 4. "This will only take a moment" (40 minutes) 5. "I'll be careful this time" @claudeai @OpenAIDevs Which one did yours say this week?

Small change on our client agents at @Arosplatforms: Every agent now ends each run with one plain sentence saying what it did and what it skipped. Not a log. A sentence a manager can read in five seconds. The skipped part turned out to be the useful half. @AnthropicAI What does your agent tell you when it finishes?

7 Codex CLI commands worth knowing before a long session, checked today in the official @OpenAIDevs docs: 1. /status shows the active model, approval policy and how much context is left 2. /compact summarizes the chat so far to free up tokens 3. /side opens a side chat that stays out of the main transcript 4. /fork copies the current chat so you can try another approach 5. /goal sets a target Codex keeps tracking, up to 4,000 characters 6. /diff shows the Git diff, including files Git isn't tracking yet 7. codex exec --worktree runs a task in its own managed Git worktree What's the first command you run in a new session?

Before a client agent at @Arosplatforms gets write access to anything, I ask three things: 1. If it runs twice, what breaks 2. Who notices, and how fast 3. Can we undo it without calling the vendor Any "not sure" and it stays read only for another week. Same three whether it runs on @OpenAIDevs or @claudeai models. How long do you keep a new agent on read only?

Rule we use before any client agent goes live at @Arosplatforms: Run it on last month's real emails first, not the clean examples the client wrote for us. The examples are always tidy. The real inbox has forwards of forwards, half filled forms and someone replying "see below" with nothing below. If it can't handle last month, it isn't ready for next month. @OpenAIDevs What's the messiest real input your agent has had to handle?