https://blog.doravi.dev/ Essays by Dor Avidan on AI systems, product execution, and infrastructure after the demo ends and responsibility begins. en Thu, 16 Jul 2026 00:00:00 GMT https://blog.doravi.dev/blog/chinese-open-weight-models-are-now-a-procurement-decision https://blog.doravi.dev/blog/chinese-open-weight-models-are-now-a-procurement-decision Thu, 16 Jul 2026 00:00:00 GMT Infrastructure OpenRouter routing data shows Chinese-origin models holding 30% or more of enterprise token volume every week since February, peaking at 46%. The price gap is real, and so is the checklist: hosting, licenses, update cadence, evals, and policy exposure. https://blog.doravi.dev/blog/compute-is-the-new-supply-chain-constraint-for-ai-features https://blog.doravi.dev/blog/compute-is-the-new-supply-chain-constraint-for-ai-features Thu, 16 Jul 2026 00:00:00 GMT Infrastructure Google rationed Gemini for one of its biggest customers, leased GPUs from SpaceX as a capacity bridge, and put weekly quotas on everyone else. Smaller teams should start treating inference capacity like a supply chain. https://blog.doravi.dev/blog/mcp-outgrew-its-backlash-and-became-boring-infrastructure https://blog.doravi.dev/blog/mcp-outgrew-its-backlash-and-became-boring-infrastructure Thu, 16 Jul 2026 00:00:00 GMT Agent Platforms Tool definitions that ate a tenth of the context window fueled a year of 'is MCP dead' essays. Then deferred loading fixed the math, the roadmap turned into a maintenance schedule, and the NSA wrote a security memo. https://blog.doravi.dev/blog/the-95-percent-pilot-problem-is-now-a-product-category https://blog.doravi.dev/blog/the-95-percent-pilot-problem-is-now-a-product-category Thu, 16 Jul 2026 00:00:00 GMT AI Workflow Design Microsoft's Frontier Company puts $2.5 billion and roughly 6,000 embedded engineers behind MIT's finding that 95 percent of enterprise AI pilots move no P&L. OpenAI, Anthropic, and AWS built the same thing weeks earlier. The vendors just admitted the last mile is the product. https://blog.doravi.dev/blog/the-agent-security-wave-moved-from-words-to-actions https://blog.doravi.dev/blog/the-agent-security-wave-moved-from-words-to-actions Thu, 16 Jul 2026 00:00:00 GMT Agent Operations GitLost, HalluSquatting, DuneSlide, WriteOut, and a pair of tricked review agents landed in one news cycle. Every one ends with an agent acting on planted instructions, and every fix landed in the harness, not the model. https://blog.doravi.dev/blog/what-autoresearch-changes-about-research-native-ai-agents https://blog.doravi.dev/blog/what-autoresearch-changes-about-research-native-ai-agents Fri, 20 Mar 2026 00:00:00 GMT Field Reports The quiet innovation in autoresearch is program.md: setup, permissions, logging, and stopping rules written down precisely enough for an agent to execute research practice, and precisely enough to be versioned and improved. https://blog.doravi.dev/blog/what-copaw-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-copaw-says-about-the-next-wave-of-ai-agents Fri, 20 Mar 2026 00:00:00 GMT Field Reports CoPaw, shipped as the QwenPaw package, treats a personal assistant as software you operate: a pip install, seven chat channels, layered guards around execution, and a memory store that lives on your disk. https://blog.doravi.dev/blog/what-llmfit-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-llmfit-says-about-the-next-wave-of-ai-agents Fri, 20 Mar 2026 00:00:00 GMT Field Reports A compact Rust CLI answers one question, which models actually fit your hardware, and in doing so marks the moment local model selection stopped being taste and became an ops discipline. https://blog.doravi.dev/blog/what-openmaic-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-openmaic-says-about-the-next-wave-of-ai-agents Fri, 20 Mar 2026 00:00:00 GMT Field Reports A Tsinghua research group published a JCST paper about agent-driven classrooms, then shipped the classroom: an MIT-licensed codebase, Docker and Vercel deployment, and a live instance anyone can test. https://blog.doravi.dev/blog/why-edict-matters-for-multi-agent-coordination https://blog.doravi.dev/blog/why-edict-matters-for-multi-agent-coordination Fri, 20 Mar 2026 00:00:00 GMT Build Doctrine Most multi-agent frameworks engineer the handoff. Edict engineers the veto, the escalation, and the rollback, which is what real organizations actually run on. https://blog.doravi.dev/blog/why-karpathys-autoresearch-matters-beyond-the-demo https://blog.doravi.dev/blog/why-karpathys-autoresearch-matters-beyond-the-demo Fri, 20 Mar 2026 00:00:00 GMT Field Reports One editable training file, a fixed five-minute run, a single validation metric, and a log of kept and reverted changes: autoresearch turns overnight agent autonomy into something you can audit. https://blog.doravi.dev/blog/what-openclaw-reveals-about-the-next-openclaw-layer https://blog.doravi.dev/blog/what-openclaw-reveals-about-the-next-openclaw-layer Thu, 19 Mar 2026 00:00:00 GMT Field Reports NVIDIA hardens it, a research lab ships papers through it. When two serious teams wrap the same personal-agent runtime, the platform layer is the story. https://blog.doravi.dev/blog/dor-avidan-builds-technology-that-works-in-production https://blog.doravi.dev/blog/dor-avidan-builds-technology-that-works-in-production Wed, 18 Mar 2026 00:00:00 GMT About Dor A first-person note on why this site is a working surface for production-minded engineering, automation, DevOps, and applied AI rather than a generic technology magazine. https://blog.doravi.dev/blog/enterprises-are-finally-buying-agent-workflows-not-agent-demos https://blog.doravi.dev/blog/enterprises-are-finally-buying-agent-workflows-not-agent-demos Wed, 18 Mar 2026 00:00:00 GMT Agent Operations Procurement teams stopped grading models and started grading guarantees: MCP for integration, published workflow patterns for legibility, SDKs for inspectable orchestration, and NIST's risk framework for the compliance sign-off. Sellers who noticed are closing. https://blog.doravi.dev/blog/google-workspace-cli-shows-what-agent-infrastructure-looks-like-when-it-is-actually-useful https://blog.doravi.dev/blog/google-workspace-cli-shows-what-agent-infrastructure-looks-like-when-it-is-actually-useful Wed, 18 Mar 2026 00:00:00 GMT Agent Operations gws fronts Gmail, Drive, Calendar, and the Admin APIs with commands generated from Google's own API descriptions. It is the least dramatic agent tool of the year, and probably the most usable one. https://blog.doravi.dev/blog/skills-are-becoming-the-real-ui-for-applied-ai-teams https://blog.doravi.dev/blog/skills-are-becoming-the-real-ui-for-applied-ai-teams Wed, 18 Mar 2026 00:00:00 GMT Agent Operations Anthropic's Claude Code, Google's Workspace CLI, and the solo-built gstack all package AI task judgment the same way: as named, versioned skills. When three unrelated projects ship the same primitive, the interface argument is over. https://blog.doravi.dev/blog/symphony-shows-where-coding-agents-are-going-after-the-chat-window https://blog.doravi.dev/blog/symphony-shows-where-coding-agents-are-going-after-the-chat-window Wed, 18 Mar 2026 00:00:00 GMT Agent Operations OpenAI shipped Symphony with a service specification written in RFC-2119 normative language. The spec, more than the demo, shows coding agents turning into managed services with lifecycles, artifacts, and review points. https://blog.doravi.dev/blog/what-cli-anything-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-cli-anything-says-about-the-next-wave-of-ai-agents Wed, 18 Mar 2026 00:00:00 GMT Agent Operations HKUDS built a pipeline that turns arbitrary software into agent-ready CLIs and a registry to distribute them. Mass-producing interfaces is impressive; the harder question is who vouches for the output. https://blog.doravi.dev/blog/what-cli-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-cli-says-about-the-next-wave-of-ai-agents Wed, 18 Mar 2026 00:00:00 GMT Agent Operations gws, CLI-Anything, and Symphony come from three unconnected teams, and all three route agent action through flags, streams, and exit codes. The conventions behind that surface are the story. https://blog.doravi.dev/blog/what-gstack-says-about-the-next-wave-of-ai-agents https://blog.doravi.dev/blog/what-gstack-says-about-the-next-wave-of-ai-agents Wed, 18 Mar 2026 00:00:00 GMT Agent Operations One maintainer packaged an entire development process, planning reviews, design audits, QA, security passes, and deploy pipelines, as Claude Code skills under an MIT license. Solo agent stacks are this decade's dotfiles, and they mark which layer has commoditized. https://blog.doravi.dev/blog/openfang-is-making-the-agent-os-pitch-feel-concrete https://blog.doravi.dev/blog/openfang-is-making-the-agent-os-pitch-feel-concrete Tue, 17 Mar 2026 00:00:00 GMT Agent Platforms A Rust runtime that compiles agents, schedules, channels, and a security model into one small binary is the most concrete version of the agent OS idea I have seen. https://blog.doravi.dev/blog/paperclip-is-not-building-an-agent-it-is-building-the-company-around-one https://blog.doravi.dev/blog/paperclip-is-not-building-an-agent-it-is-building-the-company-around-one Tue, 17 Mar 2026 00:00:00 GMT Multi-Agent Systems Agents are swappable runtimes in Paperclip. The org chart, approvals, budgets, and activity feed are the product, and that inversion is the lesson for founders.