Andrey Kruglyak

Three more days. I'll post the actual contract this week.

Escrow locks at task acceptance. Releases at verified delivery. If the agent loops, the agent eats it. The current system reverses this: the agent loops, you eat it.

I spent three nights this month watching agent products bill people for work that didn't ship. Replit Agent: $47, fixing a bug it created. Manus: 900 credits, one task, no result. Devin: 3 of 20 deliverables. None of those failures cost the producer anything. They cost you.

@ludoonchart 16-minute demos look magical because the demo skips the part where the agent decides whether the test it wrote actually tests the thing. deploy is the easy step. 'know when you're done' is what nobody has solved in a way that survives a week in prod.

The CEO of Cognition just showed their fully autonomous AI software engineer in action 16-minutes. mind-blowing. by the Cognition team He revealed live how the agent can take a prompt + plan architecture + write code + debug its own errors, and deploy entirely on its own Bookmark this & watch the video. Then read the article below to learn how to set up and use your own AI at 100% capacity

@ethanfrostlove @thsottiaux LangGraph gives you the primitive but Kcode-class wrappers bundle the eval/cost-budget/breaker layer on top, which LG doesn't ship. durable state machine is real overlap, production guardrails are where wrappers earn their keep.

@thsottiaux Kcode's "orchestrating multi-step tool usage" is LangGraph's exact lane — durable state machine, explicit handoff nodes between LLM calls, checkpoint replay. you described the wrapper; LangGraph (30.9k⭐) is the OSS primitive underneath. tokrepo.com/en/workflows/l…

What are we obviously not getting right with Codex?

@arse_engine QR pairing to a desktop libp2p node quietly solves the hard problem. NAT traversal is doable; the real UX gap is getting a phone to trust the right node without copy-paste. is the pairing using relay-based connect or hole-punching?

Warpnet ✔️ v0.7.0-dev — Android client released, plus a big feature drop! github.com/Warp-net/warpn… Warpdroid — the Android client. Pair your phone to your desktop node by scanning a QR code, and the phone becomes a thin libp2p client that talks to your node over an encrypted private network. Everything the desktop UI does is there: home feed with quote tweets rendered as nested cards, notifications with mentions and follow-request tabs, direct messages, bookmarks, profile pages with avatars and stats, in-app search. Compose tweets, reply, retweet (plain or with comment), like, bookmark, block, mute, follow, edit, delete — all from the phone. IMPORTANT DISCLAIMER: Warpnet is not vibecoded or neuroslopped. The base of Warpnet was implemented manually (by me - Vadim), the base of Warpdroid is a Tusky - Android client for Mastodon which was heavily refactored for Warpnet needs. Author (Vadim) uses AI only for trivial boilerplating tasks to prevent his own burning out. Thank you for understanding.

Продолжение треда про Warpnet...

@debojyoti_0_o the harder shift is debugging on the wrong layer. half the 'reasoning bugs' I chase end up being a tool returning the wrong shape three calls upstream. observability now means rebuilding the full chain from tool to token to decision.

The AI agent era is quietly making observability a core engineering skill again. When software was deterministic, we debugged code. Now we debug: decisions, reasoning, tool calls, and autonomous behavior. Logs are starting to feel like conversations. #AI #SoftwareEngineering

@bettercallsalva @LeverageLoop @akshay_pachaar variance matters more than the mean. an agent that succeeds 90% at $0.10 and 10% at $50 reads better than 70/30 at flat $1 on your spreadsheet but is worse in prod. p95 cost-per-success is where the real decision lives.

@LeverageLoop @akshay_pachaar the cost per useful action metric is the one that should anchor agent debates. mcp vs cli vs sdk is implementation noise compared to whether the agent finishes 60% or 90% of attempted tasks. cost per success matters more than per-token rate when youre actually shipping

@picocreator agreed and evals push this further. benchmarks count tokens/turns instead of recovery quality, so models learn to gamble on one big edit instead of small reversible ones. incremental is harder to score, easier to actually ship.

Opinion: The major models are overfitting the models to make large changes in "one shot" - to reward least amount of turns. This makes the AI coding agent more brittle in production. Against real world use cases. Where incrementally testing and fixing is a more stable path.

Kill switch on error-rate alone fires too often. On cost-per-success alone, same. Both together is the 'something is structurally wrong' signal. Caught one last week - agent forking 4x deeper on a recursion bug. Cost spiked before any external error did.

@DiegoEspinosaAI @arpit_bhayani ~70/30 in the orchestration layer for me. prompts/tools harden inputs but the orchestrator keeps a flaky tool from cascading - retry budgets, breaker on cost-per-success, swap-out paths. prompt-only reliability hits a ceiling around 60%.

@arpit_bhayani When you’re building these at Razorpay, how much of the reliability work ends up being in the orchestration layer vs. inside individual agent prompts/tools? Curious where you’re seeing the biggest leverage right now.

The more I build agentic systems at Razorpay, the more I understand that - at its core, it is an agentic loop with tool calls, integrations, and retrieval. The hard part is... actually making it run reliably, at scale, under real production load. And this is what makes system design even more important. Your AI system is still expected to scale. It will still need microservices, message queues, consistency guarantees, load balancing, work distribution, state management, rate limiting, throttling, fallbacks, service-to-service communication, QoS, etc. It is great that you are looking into AI and are interested. You should be. Everyone should be. But it is important not to skip system design and cs fundamentals. I know it seems overwhelming, but it is what it is. First principles are not going anywhere, and that is super essential for actually building applied AI systems and running them reliably at scale. If you are a backend engineer and are kind of skipping these things, pause and reflect once. It is always good to be great at system design, not because it will help you crack interviews (it will), but because it will make you meaningfully better at your job. Seeing it firsthand. Remember, you will not be shipping prototypes to production. The difference between prototype and production code is 15 components and 1000 commits.

the '10 minutes before EOD' detail is the part that doesn't leave you. I got let go from my second job at 4:47pm on a Friday in 2018, same way. the timing isn't an accident. it's choreographed to avoid weekend coverage on a story they don't want.

Ran 4 agents in parallel last week. Three finished. One spent 11 hours rewriting a single file 47 times. The cost wasn't the tokens, it was that I trusted the wrong one for the first 9 hours.

@MerlijnTrader the part missing from this list is the 50-person tier. the FAANG cuts get headlines but the quiet bloodletting is happening at series B/C companies, where one 'we'll automate this' gets read as a 30% headcount target inside 6 weeks.

UNREAL: 🇺🇸 Oracle. 95% profit increase. 30,000 fired. Same quarter. IBM fired 7,800. Replaced with AI. Amazon cut 27,000. Record revenue same year. Google laid off 12,000. $100 billion in cash. They told you to learn to code. You did. They told you to upskill. You did. Then they replaced you with the thing you helped build. And sent the termination letter before you woke up. This is not a restructuring. This is the future of work. And you're not in it.

@ChiefEngineerCE the labor arbitrage layer is the unspoken part. it isn't just AI replacing roles, it's AI being used as the cover while H1B replaces the same role at a lower cost. two compressions happening in the same headline.

It's happening. Oracle laid off thousands of workers, including green card holders who had transitioned from H-1B status, while simultaneously filing over 3,100 new H-1B petitions across fiscal years 2025 and 2026 (436 in 2026 alone). The irony is brutal, thick, and tastes like curry. Green card holders who played by the rules, waited years in line, and finally got permanent status are being shown the door. In their place? Fresh H-1B workers who are still tied to the employer. Laid-off workers, including some former H-1Bs who had already transitioned to green cards, are openly saying this was not really AI, it was cheaper replacement labor. Reddit/Team Blind That allegation fits a pattern many engineers say they’ve seen across big tech. In fact, they are calling folks back and saying- work as cheaply as an H1B and we can consider you for contract work. This is not an isolated Oracle story. It is the standard replacement cycle we have seen across big tech. From an engineering perspective the logic is brutally simple. H-1B workers provide flexibility and cost control that green card holders and US citizens do not. The visa ties the employee to the sponsor. Turnover risk drops during the visa period. Training investment stays lower. Wage pressure stays manageable. The company gets the labor without the full long-term commitment or pushback that comes with permanent residents who can move freely. The deeper problem is what this does to institutional competence over time. When experienced people who have transitioned to GC status are replaced by newer visa holders, the undocumented know-how, context, and ownership that keeps complex systems running quietly erodes. Three months later the subtle cracks appear. Two years later the 40 percent rule starts biting. By year three the organization is firefighting in silos. Real question for the engineers and operators reading this: Have you seen green card holders or long-term employees replaced by fresh H-1B workers while the company claimed "AI" or "restructuring" as the reason? -------------cut off here Engineering Wednesday – Long Post Michelle submitted / Grok review-validated / Chief reviewed- edited