Essays on AI architecture, governance, and systems that ship.https://www.langoni.me/https://www.langoni.me/writing/ai-infrastructure-stack/https://www.langoni.me/writing/ai-infrastructure-stack/A sourced map of the ten-layer AI infrastructure stack, what remains contested, and which architecture decisions European enterprises should keep reversible.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/consequential-ai-agent-infrastructure/https://www.langoni.me/writing/consequential-ai-agent-infrastructure/AI agents that affect money, data, rights, services or physical environments need bounded authority, evidence and real user recourse.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/consolidation-as-historical-norm/https://www.langoni.me/writing/consolidation-as-historical-norm/How recurring bottlenecks in radio, telephony, cable and publishing illuminate concentration in AI infrastructure, and how open institutions can change the pattern.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/day-in-the-life-2031/https://www.langoni.me/writing/day-in-the-life-2031/A grounded thought experiment about one ordinary day with delegated AI agents, digital identity, runtime controls, health support and media provenance.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/great-agi-illusion/https://www.langoni.me/writing/great-agi-illusion/AGI may transform machine capability. Access, authority, distribution and economic value will still depend on the infrastructure around it.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/how-do-you-test-a-control-that-can-never-fail/https://www.langoni.me/writing/how-do-you-test-a-control-that-can-never-fail/A practical guide to testing safety and policy controls with mutation, path evidence and an explicit UNEVALUABLE result.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/open-source-saved-the-internet/https://www.langoni.me/writing/open-source-saved-the-internet/The internet stayed contestable through open protocols, shared software and viable commercial ecosystems. AI now faces a harder version of that test.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/the-new-mainframe/https://www.langoni.me/writing/the-new-mainframe/Frontier AI has a concentrated, mainframe-shaped core. Its application edge remains more open, and that distinction will decide where power settles.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/three-ai-stacks/https://www.langoni.me/writing/three-ai-stacks/The US, EU and China are building distinct AI systems around different rules, infrastructure and economic priorities. The seams matter most.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/what-your-ai-pilot-proved/https://www.langoni.me/writing/what-your-ai-pilot-proved/An AI model can pass a pilot while the surrounding operating system fails. Production exposes integration, workflow, lineage and runtime risks the demo never tested.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/who-captures-the-rent/https://www.langoni.me/writing/who-captures-the-rent/Who captures durable value in AI infrastructure and applications? History points to the scarce complement customers cannot easily reproduce, substitute or leave.Fri, 02 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/the-model-can-be-probabilistic-the-authority-boundary-should-not-be/https://www.langoni.me/writing/the-model-can-be-probabilistic-the-authority-boundary-should-not-be/Hard rules own actions an AI agent must never take. Typed semantic judgment handles bounded ambiguity, and qualified review sees the consequential residue.Thu, 01 Oct 2026 00:00:00 GMThttps://www.langoni.me/writing/runtime-is-the-moment-of-consequence/https://www.langoni.me/writing/runtime-is-the-moment-of-consequence/The Therac-25 record shows how fatigue, unsafe assumptions and a missing effective interlock converged at execution time. Some AI-agent action paths recreate parts of that risk.Sat, 13 Jun 2026 00:00:00 GMThttps://www.langoni.me/writing/ai-governance-without-runtime-is-performative/https://www.langoni.me/writing/ai-governance-without-runtime-is-performative/Documentation matters. It becomes operational only when consequential actions pass through an enforceable boundary with reconstructable evidence.Mon, 11 May 2026 00:00:00 GMThttps://www.langoni.me/writing/ai-agent-approval-fatigue-is-real/https://www.langoni.me/writing/ai-agent-approval-fatigue-is-real/If your coding agent asks for approval on every harmless action, people either turn it off or stop using the loop. The fix is not more prompts. It is a policy layer.Sat, 11 Apr 2026 00:00:00 GMThttps://www.langoni.me/writing/the-missing-layer-in-autonomous-agents-is-runtime-policy/https://www.langoni.me/writing/the-missing-layer-in-autonomous-agents-is-runtime-policy/Teams trying to productionize autonomous agents usually do not fail because the model is weak. They fail because there is no clean policy layer between the agent and the tools it can trigger.Sat, 11 Apr 2026 00:00:00 GMThttps://www.langoni.me/writing/why-im-building-an-ai-operating-system/https://www.langoni.me/writing/why-im-building-an-ai-operating-system/Most AI tools are isolated workflows. I wanted to understand what happens when intent, execution, governance, and learning belong to the same system.Thu, 02 Apr 2026 00:00:00 GMThttps://www.langoni.me/writing/ai-code-division-of-labour/https://www.langoni.me/writing/ai-code-division-of-labour/Reliable agentic systems get faster and more truthful when models stop doing deterministic work and start handling only ambiguity.Fri, 20 Mar 2026 00:00:00 GMThttps://www.langoni.me/writing/zero-fabrication-by-architecture/https://www.langoni.me/writing/zero-fabrication-by-architecture/A model cannot alter a field it never regenerates. That protects one authoritative data path, not the factual correctness of the whole system.Sun, 08 Mar 2026 00:00:00 GMT