DEV Community: Harsh Kadyan

Tutti VM: Would You Trust a Tool to Merge Your Agents' Code?

Harsh Kadyan — Wed, 02 Sep 2026 07:33:48 +0000

hey have u guys tried Tutti VM yet

it just opened early access and the concept is wild. its a shared cloud room where ur Claude Code and ur teammates Codex can work on the same files at the same time. no copy pasting summaries back and forth no re briefing

the clever part is the agents stay local on ur own machine. ur subscription ur context ur config. only the collaboration layer goes to the cloud. so its not another hosted agent u pay for its just the glue between the ones u already use

they built their own filesystem level conflict resolution so two agents editing the same file dont blow each other up. thats the part im curious about because file conflicts are the thing that always kills parallel agent workflows in practice

free during early access. bring ur own subscription. supports Claude Code Codex Cursor and OpenCode right now

my question. would u trust another tool to sit between ur agents and handle the merging. or is that too much black box for production code

harshkadyan #thedevs #brutalharsh

Asana did 5 years of engineering in 2 weeks with 4 Codex agents

Harsh Kadyan — Tue, 01 Sep 2026 07:37:30 +0000

been using Codex and Claude Code on client projects for months now but the Asana story still hit different

they had a frontend test migration estimated at 5 years and $6M. four Codex agents did it in 2 weeks for $12K. thats not a typo. twelve thousand dollars

the workflow is what got me. one five sentence prompt. up to four agents running in parallel each on a separate copy of the codebase. an engineer checking in twice a day. thats it

i been doing something similar on a smaller scale. got a client project with a legacy Express backend that needs refactoring. started pointing agents at individual modules. not as dramatic as Asana but the same idea. break the big scary migration into parallel agent tasks and just review the diffs

the thing nobody talks about. the hard part isnt the AI. its writing a prompt specific enough and having tests that catch regressions. Asana had a clear finish line (remove Enzyme) and tests to verify. without that the agents just make a mess

what project have u been putting off because it felt too big? might be time to revisit it

harshkadyan #thedevs #brutalharsh

Cursor just launched Origin. a mini GitHub inside your IDE

Harsh Kadyan — Mon, 31 Aug 2026 07:31:48 +0000

hey have u guys seen what Cursor just dropped

they launched Origin. a git based code hosting platform built right inside the Cursor editor. repos, pull requests, code browsing, all without leaving the app. basically a mini GitHub inside your IDE

the timing is wild too. it went live the same day GitHub had that multi hour outage. Actions, API, git ops, Copilot all down. Cursor basically said hey if GitHub keeps breaking we got u

it syncs with GitHub so your mirrored repos stay in sync. pushes still land on GitHub. so its not replacing it yet, more like a backup layer that lives where u already work

its only on Pro, Teams and Enterprise plans. early beta. no public projects or CI yet

my question. would u move your repos to an editor owned git platform? or is that too much lock in for one company controlling your code and your editor?

harshkadyan #thedevs #brutalharsh

GPT 5.6 Luna dropped 80 percent in price so I recalculated all my API costs

Harsh Kadyan — Sun, 30 Aug 2026 07:32:31 +0000

been optimizing API costs across client projects this week

and the GPT 5.6 Luna price drop just changed everything. 80 percent cut to 20 cents per million tokens. thats insane

had 3 projects where we were routing around OpenAI because it was too expensive for high volume stuff. now reconsidering

went back and recalculated monthly burn for one client doing automated content gen. was paying around 400 a month on a cheaper model. with Luna at the new price it drops to like 80. same volume

crazy part is the quality didnt drop. Luna benchmarks are solid for what these clients need. summarization, classification, some code review stuff

spent last night swapping model names in the config and running the test suites. 2 of 3 passed clean. the third needs prompt tweaks because Luna formats things slightly differently

when a model drops 80 percent in price dont just read the news. go recalculate your actual costs. you might be surprised

harshkadyan #thedevs #brutalharsh

This anonymous AI model beat GPT 5.6 and nobody knows who made it

Harsh Kadyan — Sat, 29 Aug 2026 07:32:35 +0000

hey have u guys heard about this anonymous AI model that just showed up on OpenRouter

its called OX Alpha. nobody knows who made it. just appeared on August 20 with no company name, no branding, nothing. free to use for a week with a 1M token context window

the crazy part. people tested it on DeepSWE and it scored 80% pass rate. beating Claude at 65% and GPT 5.6 at 52%. for a model nobody can identify

community is saying its probably GLM 5.3 from Z.ai based on fingerprinting. the way it responds, the token patterns, all point to the GLM family. but nobody confirmed

i tried it through OpenCode for a couple hours yesterday. the coding is genuinely good. long context is nice for dumping whole repos. but im not putting production code through an anonymous model lol

my question. would u trust an unidentified model with your codebase? even if its free and performs well?

harshkadyan #thedevs #brutalharsh

JetBrains survey: 90% of devs now use AI coding agents, Claude Code doubled to 39%

Harsh Kadyan — Fri, 28 Aug 2026 07:36:57 +0000

JetBrains just dropped their 2026 developer survey and the numbers are actually wild

they surveyed 15,000 devs worldwide. 90% are using AI coding agents at work now. 68% daily. like its not optional anymore, its just how we build

but the part that got me. Claude Code went from 18% adoption in January to 39% now. doubled in 6 months. almost half of US devs use it. and Copilot dropped from 29% to 21% in the same window

feels accurate tbh. 6 months ago i was still doing everything manually half the time. now i basically live in terminal agents. my job shifted from writing code to reviewing code an AI wrote. and honestly the quality got good enough that sometimes i just merge after a quick scan

the scariest stat. Codex went from 3% to 16% adoption in the same 6 months. 5x growth. OpenAI is not playing around

my real question. is there a point where we stop writing code at all and just review diffs all day. because we are moving toward that faster than anyone predicted

harshkadyan #thedevs #brutalharsh

LinkedIn built a multi-agent AI that reviews 79,000 pull requests a week

Harsh Kadyan — Thu, 27 Aug 2026 07:34:41 +0000

hey have u guys seen what LinkedIn just built for code review

they made a multi agent AI system that reviews pull requests at scale. like multiple agents working together, each one focused on different things. one checks bugs, one checks security, one checks performance

already running on 10,000 repos and doing 79,000 reviews a week. and the crazy part is 63.9% of the AI comments actually got accepted by developers. thats not just noise, thats real signal

i have been using CodeRabbit on some of my repos and its decent but single agent. the multi agent approach sounds like a big upgrade because each agent has a narrower focus and less room to hallucinate

my question is, at what point does AI review replace human review entirely? or do we always need that second pair of eyes?

anyone here tried multi agent code review on their repos yet?

harshkadyan #thedevs #brutalharsh

Meta's Muse Code terminal agent keeps background agents alive

Harsh Kadyan — Wed, 26 Aug 2026 07:33:07 +0000

been trying Meta's new Muse Code terminal agent this weekend

its a coding agent that runs in your terminal but the interesting part is it keeps background agents alive during your session. like they dont start fresh for every task. they remember context and keep working in the background while u do other things

ran it on one of my repos for about 3 hours. it made like 400 tool calls, refactored a whole module, wrote tests, caught its own errors. the crash recovery is wild too. it logs every action so if it crashes it picks up where it left off

the model is called Muse Spark 1.2 and Meta trained it specifically for long repo scale tasks. in their own test it ran for 24 hours straight optimizing GPU kernels. 24 hours

my main concern is the same as with every agent tho. how much do u trust it to touch production code. i kept it on a feature branch and reviewed everything before merge

honestly feels like we are heading toward a world where u dont write code anymore. u review it. and that shift is happening faster than i thought

anyone else tried Muse Code yet or still sticking with Claude Code and Cursor?

harshkadyan #thedevs #brutalharsh

We just launched TheDevs

Harsh Kadyan — Tue, 25 Aug 2026 13:53:34 +0000

We just launched TheDevs. A software agency that actually builds things.

Not slides. Not promises. Working products.

We handle backend, web, iOS, and Android. Full stack, end to end. One team, everything shipped.

100+ projects delivered worldwide. From Discord bots to self-hosted mail servers. From B2B hotel platforms to AI agent infrastructure. Every project treated like a real product, not a throwaway script.

Too many agencies overpromise and underdeliver. We do the opposite. You tell us what you want, we build it, and it works. No fluff, no missed deadlines, no surprise invoices.

If you need a website, an API, a mobile app, an AI integration, or automation that runs in production, TheDevs can build it.

Website: https://thedevs.org
Walkthrough video: https://youtu.be/JJvDZv3l_Fg

MongoDB Atlas MCP Server lets AI coding agents talk to your live database

Harsh Kadyan — Tue, 25 Aug 2026 07:35:37 +0000

hey have u guys tried the new MongoDB Atlas MCP server

basically MongoDB made a hosted MCP server that lets your AI coding agents talk to your live database directly. Claude Code, Codex, Devin, all of them. no connection strings to paste, no infra to run.

your agent can query data, inspect schemas, update documents in real time while it writes code. 30k installs a week already

i use MongoDB in like half my projects so this is kinda huge. instead of writing queries manually or copy pasting data to the agent, it just connects and sees everything

sounds cool but also scary tbh. giving an AI write access to production data? thats a whole discussion

anyone here using MCP servers in their workflow yet? would love to hear how its going

harshkadyan #thedevs #brutalharsh

GitHub keeps going down and Cursor just launched a rival

Harsh Kadyan — Mon, 24 Aug 2026 07:34:45 +0000

been pushing some changes to production last week and GitHub decided to take a 7 hour nap

like the whole site. auth, actions, PRs, copilot, everything. just gone. 257 outages in the past year according to LeadDev. one major disruption every single week

the timing was wild because Cursor literally launched Origin the same day. their own code hosting platform. and some guy at Cursor quote tweeted the launch saying "we were going to ship this earlier but GitHub was down" lol

i have been feeling it too. half my CI runs fail because of GitHub side issues not my code. and when copilot goes down u lose ur autocomplete and ur agent runs both at once

so now there is a real question. do u start mirroring repos to Origin as a backup? or is this just another vendor lock in waiting to happen

for now i am keeping everything on GitHub but adding a second remote to Origin for my main repos. belt and suspenders

what about u guys. anyone tried Origin yet? or still riding the GitHub struggle bus?

harshkadyan #thedevs #brutalharsh

AI coding agents changed my workflow entirely

Harsh Kadyan — Sun, 23 Aug 2026 07:33:20 +0000

been using AI coding agents for a while now and honestly the workflow shift is wild

like 6 months ago i was still writing every line myself. now i basically describe what i want, review the diff, and steer it if it goes off track. not autocomplete. actual agents that read the codebase, edit multiple files, run tests, and fix their own errors

the weird part is my job is becoming more about reviewing than writing. i feel like a senior dev managing junior engineers who never sleep but will absolutely do something stupid if u stop watching for 5 min

shipped a feature last week where the agent wrote the backend, wrote the tests, caught its own bug, fixed it, and opened the PR. i just reviewed and merged. that hit different

what's ur stack looking like these days. cursor? claude code? terminal agents? genuinely curious what people are actually using day to day vs what's just hype

harshkadyan #thedevs #brutalharsh