Yaroslav

Distributed agentic deployments in big tech. Artisanal code maker, generalist skeptic. Dad, reader, humanitarian at heart. Views are my own.

[Yaroslav @yarolegovich]

@MaxHirsch13 You need two so that the development doesn't stop when one goes on vacation

[Yaroslav @yarolegovich]

@ThePrimeagen Dm for a referral :)

[Yaroslav @yarolegovich]

The core of most useful systems is a network or a marketplace, and you can't vibecode other people into it. Siloed software isn't that useful, and federated software is extremely hard.

[Yaroslav @yarolegovich]

@NickADobos Learning just in time works for the details, not the fundamentals. That's what schools teach btw

[Yaroslav @yarolegovich]

@signulll My context window is smaller but better curated

[Yaroslav @yarolegovich]

@stevejonesdev The average person won't build their own software, but companies with engineering teams might though. Rebuilding SaaS in-house used to be too expensive to justify, and this has now changed.

[Yaroslav @yarolegovich]

if everyone vibecodes their own X will they just read their own feed?

[@levelsio @levelsio]

Software is dead

It's now the era of personal software and everyone just builds what they need whenever they need it I think

[Yaroslav @yarolegovich]

@LoicBerthelot By vibecoding paid software?

[Yaroslav @yarolegovich]

@jasonlk Sounds like a "read 1,000, collect prompt tweaks, ship, repeat" infinite loop.

[Yaroslav @yarolegovich]

@bradzacher In my experience with agents it's more often about overdoing as in "why did you add that". While with humans it's about quality as in "why did you do it like that".

[Yaroslav @yarolegovich]

@poteto The best code was always the code I wrote 10 minutes ago

[Yaroslav @yarolegovich]

@lennysan @Atlassian * once engineers come on board, they get to burn the alpha and rewrite it from scratch. Possibly in Rust.

[Yaroslav @yarolegovich]

Built my wife a software factory: GitHub CI, Cursor cloud agents on the $20 plan, Expo, and a messenger bot.

She's building her startup on maternity leave by texting it ideas. The bot turns them into GitHub issues, CI drives them through implementation, testing, and deployment, and the update lands on her phone. I'm on-call.

Three years ago I'd have said this would never work.

[Yaroslav @yarolegovich]

@milan_milanovic I'd frame the choice differently. MCP is for when you're happy to expose your system building blocks and support them like an API. A2A is for when you want to expose the capabilities and keep the tools, prompts and know-how behind the interface.

[Yaroslav @yarolegovich]

@ParamSiddh it's artisanal coding

[Yaroslav @yarolegovich]

@garrytan @capydotai Agents left alone with evals for 16 hours get very good at passing the evals :)

[Yaroslav @yarolegovich]

This one's going to stay in my open tabs for a while.

On any real system you never understood most of the code. The troubleshooting loop is still the same: model the system -> check telemetry -> find a mismatch -> explain it -> adjust the model -> repeat.

Agents crunch through telemetry at insane speeds, and not using them for it is criminal. But I'm far from comfortable automating the whole loop yet.

[Yaroslav @yarolegovich]

@hwchase17 64% off the median is impressive. Curious about the rest of the distribution. How did the tail and variance move?

[Yaroslav @yarolegovich]

@rauchg Agents deploy sites, agents read them, agents file bug reports. Humans read summaries and approve the invoice. Sounds like a fun future.

[Yaroslav @yarolegovich]

@theo Coding is mostly solved. AI is still bad at asking what problem we're solving, and whether we should be solving it at all