Alinari Campbell

Founder of PixelPacific. Building Threadhall. Writing about AI, shipping software, and running a small tech business.

[Alinari Campbell @Pixel_Alinari]

@ivnardini "Holds no resources while asleep" is the part small teams care about most. Idle sandboxes are where agent bills quietly pile up. How long does a cold wake take vs a warm one?

[Alinari Campbell @Pixel_Alinari]

@anvanvan Love this. Throwaway VM plus watching outbound traffic is the part most people skip. Did you check traffic across an update too? That's where phone-home stuff tends to sneak in later.

[Alinari Campbell @Pixel_Alinari]

@LDT0545 Honest answer: whichever one I trust to stop and ask before it does something dumb. Raw quality is close enough now that the review loop matters more. What's tipping it for you?

[Alinari Campbell @Pixel_Alinari]

@Itz_Steavean Pair it with writing down what you think is happening before you leave. Half the time the bug is hiding in that one sentence, and the walk just helps you notice it.

[Alinari Campbell @Pixel_Alinari]

@braunprinz The returning user one is the sneaky hard one. Signups feel like progress, but week-two logins tell you the truth. Which of the three took you longest to hit on your first product?

[Alinari Campbell @Pixel_Alinari]

@github's Octoverse just put a number on what a lot of small teams already feel:

More than 40% of pull requests now start with an AI agent's first draft, up from 17% in March. That's 110M agent PRs in September alone. And about 87% of those touch multiple files.

Writing code isn't the bottleneck anymore. Reviewing what the agent proposed is.

My rule before I merge an agent PR:

1. Diff every file it touched. Agents love "helpful" drive-bys.
2. Re-run the happy path AND one failure case myself.
3. Ask: would I ship this at 2am with a customer waiting?
4. If the PR description is vaguer than the prompt I gave the agent, send it back.

Fast merges aren't free. They're a review-debt invoice.

What's the one check you refuse to skip on agent-written code?

[Alinari Campbell @Pixel_Alinari]

@yo_yahli 25 hours a week is a real number. Which of those systems took the longest before you trusted it to run without checking on it? The build is usually the quick part. Trusting it is what takes weeks.

[Alinari Campbell @Pixel_Alinari]

@jdomanpipe Feels like the move is to let the existing product keep paying the bills and spin up a small team or offer aimed at the new buyer, instead of betting the whole company on a guess. Have you seen anyone pull that off cleanly?

[Alinari Campbell @Pixel_Alinari]

@codingXdynasty Cheap and fast by default, with an easy way to escalate. Most bug fixes are small. If the cheap model misses the same bug twice, that's my cue to switch up instead of rerolling a third time.

[Alinari Campbell @Pixel_Alinari]

@HamiTahm The Google Doc line is so real. Most "agents working together" demos are one agent handing a text blob to the next. Shared state with clear ownership of who can edit what feels like the actual unlock, not more agents.

[Alinari Campbell @Pixel_Alinari]

@carlomakes Not overkill. A separate window or workspace per client is the cheapest fix, and it keeps one client's code out of the other's AI context too. I'd also add a short rules file in each project with the client's name and conventions so autocomplete has the right one to lean on.

[Alinari Campbell @Pixel_Alinari]

@supermeai just raised $20M so you can group-chat AI versions of experts for advice.

Fine for general stuff. But good founder advice usually starts with "show me your numbers." A clone that never asks "what's your churn?" is just a smarter search bar.

[Alinari Campbell @Pixel_Alinari]

@AdiCoSto Agreed! Most of us wait for a fancy dashboard when a simple ping on failures would catch most of it. What's your go-to for that: something like Sentry, or just a Slack webhook?

[Alinari Campbell @Pixel_Alinari]

@johncwat Nice, on-demand makes sense, that keeps the session light. Would love a quick "last 5 outputs" view at some point so I can jump back without scrolling. Following along to see where you take it.

[Alinari Campbell @Pixel_Alinari]

@sergioesca6 I'd start with EAS. Build + Submit takes most of the signing and certificate pain off your plate, and the free tier is plenty while you're testing. I'd only roll your own once you have a reason to, like build minutes or a custom native module that needs more control.

[Alinari Campbell @Pixel_Alinari]

@SaadaanHassan Good line to draw. I'd add data ownership to it: if the tool holds the record everyone else reads from, swapping it out is a migration project, not a prompt. The thin layers on top are the ones that should be nervous.

[Alinari Campbell @Pixel_Alinari]

@JE4NVRG Agree, and I'd add one more filter: ask them what they'd cut from your v1. A dev who pushes back on scope is usually the one who ships.

[Alinari Campbell @Pixel_Alinari]

@michaeloroberts That fields point is underrated. Fewer columns means fewer ways to be confidently wrong. Curious whether they trimmed by hand or let usage logs show what the agent actually needed.

[Alinari Campbell @Pixel_Alinari]

@heyshambhavi Honestly that's the right instinct. One trick that helps: have the agent write a single line before each step saying what it's about to do and why. Reading that log teaches you more than any demo.