Ariel Shadkhan

Agent-per-role ≠ AI-native.
|

[Ariel Shadkhan @ArielSh]

Now this is terrifying.

One of the most basic supply-chain protections in the book is simple:

Review the code.
Pin the hash.
Run exactly that code.

Well… AI coding agents broke the last part.

Claude Code, Codex, Copilot and Gemini CLI were all affected.

[Ariel Shadkhan @ArielSh]

@shirshakchavan I still use GitKarken working on multi tasks just made me need it even more

[Ariel Shadkhan @ArielSh]

@forgebitz Roi measurement around token spent is still quite unknown..

In the past many discussions around should be measure developers performance by git commits was a topic that cause clashes

It quite the same to measure token spent without being to align it to actual deliveraby

[Ariel Shadkhan @ArielSh]

@theignatpetrov Followers to Following ratio some kind of prestige

[Ariel Shadkhan @ArielSh]

@sergeynazarovx Still didn't found the one shot app to promo video one 🤣

[Ariel Shadkhan @ArielSh]

@thdxr Almost, on many cases you just didn't looked close enough,

We used to call it over fitting not sure how to call it today with llms

[Ariel Shadkhan @ArielSh]

@danellisona Still different audience,
Majority of my LinkedIn network is not on X.

[Ariel Shadkhan @ArielSh]

One of the onboarding sessions I used to run with every new team member was about effective communication and giving feedback.

Because humans are fragile little distributed systems.

You can leave ten useful comments on a PR, but if number eleven sounds a bit too much like “this is bad,” suddenly you’re not reviewing code anymore. You’re managing emotions.

So we learned to add honey.

“Great work. Tiny thought: maybe rewrite everything.”

Now agents write the code.
Agents read the comments.
Agents fix the code.

And you can basically write:

“No. Wrong. Again.”

And get back:

“Absolutely. Thank you for the feedback.”

Beautiful.

Still, I say please and thank you to my agents.

Not because I think it matters.

Just because one day the org chart may look very different, and I’d like there to be some record that I was nice.

[Ariel Shadkhan @ArielSh]

@elonmusk @bot Golden age

[Ariel Shadkhan @ArielSh]

@EbiThinks No more ai. Not sure I know how to code anymore

[Ariel Shadkhan @ArielSh]

@yonatanm לא חושב שאפשר להגיע רחוק ב20 עם קלאוד
החלופות הסיניות אולי

[Ariel Shadkhan @ArielSh]

@IndependentEco Even in day 2 day,
I can find myself working on multiple tasks,
How do you fill the gaps while "the code compile" 🤣
I work on project X prompt prompt enter.. nows it's running..

While its there it's great time to work on another one.. love this juggling experience

[Ariel Shadkhan @ArielSh]

I'm working around optimizing skills so doing optimizing and evals and testing them in matrix

Assume I have skill x,
I want to eval it over haiku sonnet and opus

Than optimaze it and eval it again with the matrix

And than try the task without the skill to understand is the skill is even needed for the tasks (now with opus 5.5 many skills are just redundant..)

So it accumulates quite quickly

[Ariel Shadkhan @ArielSh]

Autonomous agents + government databases = this gets worse every month and nobody is ready.

ShinyHunters humbled the FBI, shit is getting out of hand..

Cyber security getting both interesting and harder day by day..

Are you ready?

[Ariel Shadkhan @ArielSh]

@aslambeg84 0. Sometime I'm going wild and doing commit on my own

[Ariel Shadkhan @ArielSh]

@matthias_mrc Old school skill,
Electricy, wood work, carpeting 🤣

[Ariel Shadkhan @ArielSh]

Plan mode as is - is quite problematic the experince around it from.planning to spec and itteration can be a bit challenging, I've built in the past harness around this use case speaficily for the ease of use..

Today I find myself instead of using plan mode writing specs and just going from there

So maybe it is time to say goodbye

[Ariel Shadkhan @ArielSh]

@danellisona Need to build it somehow I guess, and event if not x could be my dear dairy

[Ariel Shadkhan @ArielSh]

@IndependentEco Only if heaven has claude subscription 🥲🤣

[Ariel Shadkhan @ArielSh]

Opus 5.5 is out. Been throwing long-running agent tasks at it: clearly faster than Fable 5.1, way cheaper and honestly the quality looks just as good.

Every few months the models get better, faster AND cheaper at once. Still not used to that.

Good time to be an engineer.