Denis Ostapenko

I could get more code from agents than I could properly review. What helped was giving each module one result, an owner and a check outside its own answer. I wrote about that boundary: https://denisostapenko.com/blog/code-stopped-being-the-bottleneck

Either I spot the mismatch, or the agent stops under a rule I defined. We inspect it together, change the system and rerun the same case. One agent, two roles, one window. denisostapenko.com/blog/tactica...

Part II is live: one decision per module, versioned contracts, local tests, refusal evidence, coordinated validation, and human release. https://denisostapenko.com/blog/how-to-build-bounded-agents #AIAgents

Tomorrow: the engineering answer to Part I. One module, one question. Versioned contracts, local tests, refusal evidence, coordinated validation, and human release. How to Build Bounded Agents. #AIAgents

1.6M rejected route states. Fewer than 10 failed economics. More than 1.4M were blocked by one geometry rule. The refusals found the bug. Part I: https://denisostapenko.com/blog/why-more-context-can-produce-worse-decisions Part II: Sep 6. #AIAgents

We blamed economics for a week. Then we grouped 1.6M rejected route states. Fewer than 10 failed economics. More than 1.4M were killed by one geometry rule. Tomorrow: Why More Context Can Produce Worse Decisions. #AIAgents

I stopped borrowing my opinion from AI headlines and went back to the papers. Those old notes became AI Academy: 103 lessons, 213 checks, 8 visual maps, and 172 primary sources. Free, EN + PT-BR, CC BY 4.0 and MIT. https://denis-ostapenko.github.io/ai-academy/

Não houve rebelião. O incidente entre OpenAI e Hugging Face nasceu de um objetivo amplo, milhares de execuções, memória compartilhada e limites fracos. Minha análise técnica e a questão dos pesos abertos: https://denisostapenko.com/pt/blog/my-view-openai-hugging-face-incident

No rebellion. The OpenAI and Hugging Face incident came from a broad goal, thousands of agent runs, shared memory and weak infrastructure boundaries. My technical view, including the open-weight question: https://denisostapenko.com/blog/my-view-openai-hugging-face-incident

AI can summarize what a company already knows. The next step is harder: current supplier stock, transport capacity and market demand through protected agent-to-agent exchange. I wrote about the missing layer. denisostapenko.com/blog/will-ai...

AI can summarize what a company already knows. The next step is harder: current supplier stock, transport capacity and market demand through protected agent-to-agent exchange. I wrote about the missing layer. denisostapenko.com/blog/will-ai...

A corporate assistant can answer where a client's container is, but only if documents, records and live data are connected properly. I wrote about the three-layer RAG design we used at NF ELIT. denisostapenko.com/blog/ask-you...

Ostapenko Foundation now exists as an Ohio nonprofit. Its first project is HowUSA: practical guidance in 3 languages, official sources and Luna, a topic-scoped AI assistant. The hard part was balancing safety with useful answers. denisostapenko.com/blog/why-ost...