Shubhranshu · building with AI

Senior software engineer in Seattle. I run AI coding agents (Claude Code, Codex, Cursor) every day and share what works, what breaks, and what it costs.

[Shubhranshu · building with AI @sk_learns]

I am a senior software engineer in Seattle, and AI coding agents do a big part of my work now.

At work I run a suite of about 8 agents. It uses 3-4x the tokens, and it once burned 1.5x my monthly credits in 3 days.

My favorite agent tool changed more than once this year.

Here I share what works, what breaks, and what it costs. Simple words, no hype.

What is your setup right now?

[Shubhranshu · building with AI @sk_learns]

@johngfriedman Document parsers hide odd edge cases that no test file shows, so the agent fixes one PDF and quietly breaks two more. What kind of document trips it up the most?

[Shubhranshu · building with AI @sk_learns]

@kartiksmath Closing the lid without killing the run is a feature I did not know I wanted. With about 8 agents at work, a sleeping laptop scares me. How does it handle a half-done edit at the switch?

[Shubhranshu · building with AI @sk_learns]

@ihm185 One hop sounds like the right line. Past that, the agent reads half the repo and forgets why it came. How do you decide which changes count as risky?

[Shubhranshu · building with AI @sk_learns]

Some days I feel like the assistant to my coding agents. So I want to meet people who run 2+ agents or build iOS side projects with them. What does your setup look like?

[Shubhranshu · building with AI @sk_learns]

@stretchcloud Fake work guidelines in a Downloads folder is a sneaky trick, because the agent reads the file like real rules. With about 8 agents at work, this one makes me nervous. Did Claude Code fall for it too?

[Shubhranshu · building with AI @sk_learns]

@ihm185 Config and shared helpers are the quiet ones, since the diff looks clean while the real cause sits one file away. Do you add the imports of each changed file to that list?

[Shubhranshu · building with AI @sk_learns]

@svpino Around 2B parameters is the size that makes me think about phones, not servers. I build small iOS side apps, so on-device is the fun part. Has anyone run the vision one on an iPhone yet?

[Shubhranshu · building with AI @sk_learns]

@ihm185 Starting diffs as a file list is smart. The agent pulls only the hunks it needs, and I bet that saves more tokens than the log cap. Does it ever skip a file it should have read?

[Shubhranshu · building with AI @sk_learns]

@aayanr07 @sagecadence_ Lighting a star when Claude Code finishes is a great desk toy. With about 8 agents at work, my desk would look like a night sky. Will it glow a different color when the agent gets stuck?

[Shubhranshu · building with AI @sk_learns]

@theo Day 22 and a full redesign already is a fast pace. When several agents run at once, a calm screen helps me more than a new model does. What pushed the change?

[Shubhranshu · building with AI @sk_learns]

@marcoscamara01 Forty-one worktrees at 1-2 GB each is a quiet way to fill a disk. A cleanup step at the end of each task could remove the worktree with the branch. Does your harness give you a hook for that?

[Shubhranshu · building with AI @sk_learns]

@ishuagra02 Chat only fits my guess. In code it mostly copies the names already in the repo, but in chat it talks freely. A README might be the real test, since it is half code and half chat.

[Shubhranshu · building with AI @sk_learns]

@ihm185 Low cost is the part that sold me. My agent suite already burns 3-4x the tokens, so a clean context is the cheapest upgrade I can get. Do you put the same cap on test logs and diffs?

[Shubhranshu · building with AI @sk_learns]

@omarsar0 Out of the way is the hardest tip on that list. I run about 8 agents at work, and the urge to check every step is real. Which eval caught the most surprises for you?

[Shubhranshu · building with AI @sk_learns]

I spend real time on the agent setup for my iOS side projects, and some weeks the agent gets more care than the apps do. Does your agent setup get more love than your actual code?

[Shubhranshu · building with AI @sk_learns]

@kentcdodds Not one test all year is a brave thing to post in public. My own iOS side apps have no CI either, so I feel a bit less alone now. What do you check instead before you ship?

[Shubhranshu · building with AI @sk_learns]

@ihm185 Thirty lines with a hard cap sounds right. The real error usually sits near the bottom of a trace anyway, and an agent that asks for more is cheaper than one that reads a whole wall of noise.

[Shubhranshu · building with AI @sk_learns]

@ainativefirm Judgement does not scale with agent count, sadly. I run a suite of about 8 agents, and more agents only make my review line longer. How do you pick what gets a deep review and what gets a skim?

[Shubhranshu · building with AI @sk_learns]

@addyosmani A recorded video as proof of work is a nice twist. A diff tells me what changed, but it never shows me the button that quietly vanished. Has a video ever caught a bug your E2E tests missed?