Kirill Lutcenko
@lutkir.bsky.social
AI Engineer & Back-End Technical Lead. MSc AI.
10y building software, 2y in AI and tech lead roles. Engineering-first views on what works in AI (and what doesn’t).
Blog: lutkir.dev
10y building software, 2y in AI and tech lead roles. Engineering-first views on what works in AI (and what doesn’t).
Blog: lutkir.dev
Looking for advice.
We’re having a discussion in the team right now.
We have an agent that’s supposed to perform similar, but slightly different tasks. The exact strategy depends on the task type.
1/2
We’re having a discussion in the team right now.
We have an agent that’s supposed to perform similar, but slightly different tasks. The exact strategy depends on the task type.
1/2
January 20, 2026 at 11:32 AM
Looking for advice.
We’re having a discussion in the team right now.
We have an agent that’s supposed to perform similar, but slightly different tasks. The exact strategy depends on the task type.
1/2
We’re having a discussion in the team right now.
We have an agent that’s supposed to perform similar, but slightly different tasks. The exact strategy depends on the task type.
1/2
My team as of today
January 19, 2026 at 9:53 PM
My team as of today
TIL about rclone. What a joy that such projects exist and that cloud providers support them.
January 19, 2026 at 9:17 PM
TIL about rclone. What a joy that such projects exist and that cloud providers support them.
I don't like the ambiguity of the word "Agent" in our professional lexicon.
We call a standalone executable running in a separate process an "agent."
We call a standalone executable running in a separate process an "agent."
January 18, 2026 at 8:13 PM
I don't like the ambiguity of the word "Agent" in our professional lexicon.
We call a standalone executable running in a separate process an "agent."
We call a standalone executable running in a separate process an "agent."
I think the industry now desperately needs some kind of “agent transparency and observability standard.” Something like mcp/a2a, but not for rcp, rather for interpreting agent behavior.
1/4
1/4
what's the best way to run many agents, with local issue tracking (i like beads for this) but have oversight everywhere. Think ~3-5 agents running "on cloud", and 1-2 agents on my mac, but i can see what any/all of them are doing and see if anything needs kicking.
January 18, 2026 at 7:53 PM
I think the industry now desperately needs some kind of “agent transparency and observability standard.” Something like mcp/a2a, but not for rcp, rather for interpreting agent behavior.
1/4
1/4
Just shipped my first blog post, sharing lessons learned while developing an agent and an MCP server for log-based issue investigation.
Written by a human 🙂
Feedback welcome.
Written by a human 🙂
Feedback welcome.
Log-Based Issue Investigation with AI Agents: Designing an MCP That Actually Works
I believe that, despite the multimodality of modern models, the areas where LLM-based solutions are truly useful and 100% worth investing in are those where you need to extract insights from textual d...
www.lutkir.dev
January 17, 2026 at 4:15 PM
Just shipped my first blog post, sharing lessons learned while developing an agent and an MCP server for log-based issue investigation.
Written by a human 🙂
Feedback welcome.
Written by a human 🙂
Feedback welcome.
As a dedicated @proton.me business user for several years, I decided to give their AI assistant Lumo a try. It’s actually not bad at all. It’s certainly not (and won’t be) as capable as ChatGPT or Gemini, but for simple tasks like text summarization and grammar fixes it works just fine.
January 17, 2026 at 10:05 AM
As a dedicated @proton.me business user for several years, I decided to give their AI assistant Lumo a try. It’s actually not bad at all. It’s certainly not (and won’t be) as capable as ChatGPT or Gemini, but for simple tasks like text summarization and grammar fixes it works just fine.
I'm surprised that many MCP docs/courses emphasize stdout transport over streamable HTTPS so heavily. Does anyone even use stdout as a transport of choice for agents in production?
January 15, 2026 at 1:02 PM
I'm surprised that many MCP docs/courses emphasize stdout transport over streamable HTTPS so heavily. Does anyone even use stdout as a transport of choice for agents in production?
LinkedIn is such a doomed place. It’s just a bunch of AI generated slop
January 15, 2026 at 11:57 AM
LinkedIn is such a doomed place. It’s just a bunch of AI generated slop
+ cost monitoring and assessment of newly released models
in case you've missed it, here are some G𝗼𝗹𝗱-𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱𝘀 𝘁𝗼𝗼𝗹𝘀 for AI pipelines:
A/B testing
randomized controlled trials
natural experiments.
as well as 𝘱𝘰𝘴𝘵-𝘥𝘦𝘱𝘭𝘰𝘺𝘮𝘦𝘯𝘵 𝘮𝘰𝘯𝘪𝘵𝘰𝘳𝘪𝘯𝘨
A/B testing
randomized controlled trials
natural experiments.
as well as 𝘱𝘰𝘴𝘵-𝘥𝘦𝘱𝘭𝘰𝘺𝘮𝘦𝘯𝘵 𝘮𝘰𝘯𝘪𝘵𝘰𝘳𝘪𝘯𝘨
January 13, 2026 at 7:47 PM
+ cost monitoring and assessment of newly released models
Reposted by Kirill Lutcenko
When are impossibility proofs misleading? In infinitefaculty.substack.com/p/be-wary-of..., I discuss a common issue I see: proofs that are logically valid, but where the underlying assumptions are unjustified. I discuss ‘proofs’ that cognition cannot be tractably learned, and that LMs are 1/
Be wary of assumptions in impossibility arguments
A proof is only as good as its assumptions
infinitefaculty.substack.com
January 13, 2026 at 3:33 PM
When are impossibility proofs misleading? In infinitefaculty.substack.com/p/be-wary-of..., I discuss a common issue I see: proofs that are logically valid, but where the underlying assumptions are unjustified. I discuss ‘proofs’ that cognition cannot be tractably learned, and that LMs are 1/
I think it’s the right direction for Apple.
1. They quit a race they had already lost and are concentrating on things they’re actually strong at.
2. They rightly chose Google over OpenAI, because OpenAI can’t be a reliable long-term partner given its catastrophic financial situation.
Good luck.
1. They quit a race they had already lost and are concentrating on things they’re actually strong at.
2. They rightly chose Google over OpenAI, because OpenAI can’t be a reliable long-term partner given its catastrophic financial situation.
Good luck.
Joint statement from Google and Apple
Apple and Google have entered into a multi-year collaboration under which the next generation of Apple Foundation Models will be based on Google's Gemini models and clou…
blog.google
January 13, 2026 at 3:02 PM
I think it’s the right direction for Apple.
1. They quit a race they had already lost and are concentrating on things they’re actually strong at.
2. They rightly chose Google over OpenAI, because OpenAI can’t be a reliable long-term partner given its catastrophic financial situation.
Good luck.
1. They quit a race they had already lost and are concentrating on things they’re actually strong at.
2. They rightly chose Google over OpenAI, because OpenAI can’t be a reliable long-term partner given its catastrophic financial situation.
Good luck.
Same :(
Just created this account a few days ago, subscribed to hundreds of AI and tech leaders, and suddenly fell into the rabbit hole of all the horrors American society is going through right now.
Sending you all my hugs and thoughts from the other side of the ocean 🫂
Just created this account a few days ago, subscribed to hundreds of AI and tech leaders, and suddenly fell into the rabbit hole of all the horrors American society is going through right now.
Sending you all my hugs and thoughts from the other side of the ocean 🫂
I should start watching less of these ICE videos on Bluesky. I am not American, I can't do anything about it, and it starts affecting my mental health. So much unnecessary suffering.
January 12, 2026 at 5:32 PM
Same :(
Just created this account a few days ago, subscribed to hundreds of AI and tech leaders, and suddenly fell into the rabbit hole of all the horrors American society is going through right now.
Sending you all my hugs and thoughts from the other side of the ocean 🫂
Just created this account a few days ago, subscribed to hundreds of AI and tech leaders, and suddenly fell into the rabbit hole of all the horrors American society is going through right now.
Sending you all my hugs and thoughts from the other side of the ocean 🫂
Just created the most expensive arithmetic calculator ever lol. A team of four agents with two different LLMs under the hood 🤣
January 12, 2026 at 2:23 PM
Just created the most expensive arithmetic calculator ever lol. A team of four agents with two different LLMs under the hood 🤣
True. Ten years in the industry, and I still have no idea how (or if) it’s possible to measure engineering productivity solely using hard numeric metrics, without human manager feedback.
Ppl forget that AI doesn’t change some basics:
1. Evaluate on a single metric and engineers game it (we are smart enough)
2. Code is increasingly promoted by AI
3. The single best eng contribution can be… not shipping code!
4. A 1-character change can have massive impact
Etc etc
1. Evaluate on a single metric and engineers game it (we are smart enough)
2. Code is increasingly promoted by AI
3. The single best eng contribution can be… not shipping code!
4. A 1-character change can have massive impact
Etc etc
January 12, 2026 at 2:05 PM
True. Ten years in the industry, and I still have no idea how (or if) it’s possible to measure engineering productivity solely using hard numeric metrics, without human manager feedback.
Playing with Google ADK this week. An amazing agentic AI framework, the best we’ve tried so far in terms of developer experience.
January 12, 2026 at 10:03 AM
Playing with Google ADK this week. An amazing agentic AI framework, the best we’ve tried so far in terms of developer experience.
"DL seems full of things that aren't grounded in theory, but are tricks/hacks just to get the damn thing to train".
So true!
So true!
Huh, that's cool. DL seems full of things that aren't grounded in theory, but are tricks/hacks just to get the damn thing to train. Is gradient clipping the same? It's been a while since I messed around with the internals of DL models.
January 12, 2026 at 8:43 AM
"DL seems full of things that aren't grounded in theory, but are tricks/hacks just to get the damn thing to train".
So true!
So true!
wow!
AI has gotten really good at theorem proving: axiommath.ai/territory/fr...
Axiom’s prover supposedly solved all 12 of 2025’s Putnam problems correctly. Source code: github.com/AxiomMath/Pu...
Axiom’s prover supposedly solved all 12 of 2025’s Putnam problems correctly. Source code: github.com/AxiomMath/Pu...
January 12, 2026 at 8:29 AM
wow!
Reposted by Kirill Lutcenko
January 11, 2026 at 10:19 AM
Always nice to get this kind of feedback from the dev team :)
January 11, 2026 at 3:11 PM
Always nice to get this kind of feedback from the dev team :)
Yes, I keep hearing about this over the last month or so, that corporations mostly use AI as an excuse for “tough decisions” rather than actually automating labor.
While there is anecdotal evidence of AI affecting layoffs, the macroeconomic data does not support this narrative fortune.com/2026/01/07/a...
AI layoffs are looking more and more like corporate fiction that's masking a darker reality, Oxford Economics suggests | Fortune
"Firms don't appear to be replacing workers with AI on a significant scale," the firm said. It suspects some are trying to "dress up layoffs" as good news.
fortune.com
January 11, 2026 at 3:03 PM
Yes, I keep hearing about this over the last month or so, that corporations mostly use AI as an excuse for “tough decisions” rather than actually automating labor.
Interesting how such copyright violations will be legally handled in cases where content was retrieved from a jailbroken LLM through an agent developed by someone not affiliated with the LLM vendor. Should we start worrying about this when developing agents for B2C products?
"In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim ... Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs."
arxiv.org/abs/2601.02671
arxiv.org/abs/2601.02671
Extracting books from production language models
Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized dat...
arxiv.org
January 10, 2026 at 7:14 PM
Interesting how such copyright violations will be legally handled in cases where content was retrieved from a jailbroken LLM through an agent developed by someone not affiliated with the LLM vendor. Should we start worrying about this when developing agents for B2C products?