{"slug": "lai-142-my-ai-setup-in-2026", "title": "LAI #142: My AI Setup in 2026", "summary": "Louis-François Bouchard, Towards AI co-founder and head of community, detailed his daily AI setup in LAI #142, running Claude Code and Codex from an always-on Mac mini while keeping reusable knowledge and skills in an Obsidian vault accessible from laptop or phone, and released the vault as an open-source template. Bouchard also announced a second book, \"AI Engineering for Production,\" launching October 20, which he said grew out of an update to \"Building LLMs for Production\" after the 2024 AI stack failed to survive 2026. The newsletter additionally highlighted Keviv's open-source Alineo toolkit, which gives coding agents isolated sandbox environments to run commands, modify files, install dependencies, and execute code, with saved and resumable sandbox state and snapshots for faster subsequent runs.", "body_md": "Good morning, AI enthusiasts!\n\nThis week, I’m opening up the AI setup I use every day: how I work across Claude Code and Codex, keep context and skills portable, run agents while I’m away from my computer, and turn repeated corrections into knowledge I can reuse. I’ve also shared the Obsidian vault behind it as an open-source template.\n\nWe also finally have some news we’ve been waiting to share: we wrote a second book. AI Engineering for Production launches October 20.\n\nPlus, this week’s reads get into:\n\nLet’s get into it!\n\nThis week in What’s AI, I’m sharing the AI setup I actually use every day. The main idea is simple: I don’t want my context, skills, or ongoing work trapped inside one agent, one chat, or one computer.\n\nI keep reusable knowledge and skills in an Obsidian vault, run Claude Code and Codex from an always-on Mac mini, and access the same workspace from my laptop or phone. I also give the agents different jobs: Claude Code handles more of the planning and coordination, while Codex often handles bounded execution and scheduled tasks.\n\nI also show how I manage usage limits, schedule nonurgent work overnight, maintain the vault as it grows, and turn repeated corrections into reusable skills instead of losing them in old conversations. The exact models and tools will change, but the part I want to keep is a portable system where knowledge survives those changes.\n\nI’ve shared the full setup, along with an open-source Obsidian vault template you can adapt, in this week’s article. [Read the full walkthrough](https://www.louisbouchard.ai/my-ai-setup/?utm_source=chatgpt.com).\n\nIf your AI gets a document question wrong, check the extracted text before changing the prompt.\n\nA PDF can look perfectly clear while the parsed version has already lost important relationships. A table value may be extracted without its column heading. A footnote may appear far from the sentence it qualifies. The words are still there, but the structure that gives them meaning is not.\n\nThis is one of those problems that is easy to underestimate in demo systems. We have an entire lesson on document parsing in our [Full Stack AI Engineering course](https://towardsai.com/academy/full-stack-ai-engineering/?utm_source=newsletter&utm_medium=email&utm_id=AItips) because, once you work with real documents, extraction quality becomes part of the system’s quality. Retrieval cannot find relationships that parsing has already broken, and a better prompt cannot reconstruct information that the model never represented correctly.\n\nA simple debugging test is to take one question the system answered incorrectly and try to answer it yourself using only the extracted text. If the answer is missing, ambiguous, or difficult to reconstruct from that version, fix the parsing first. Otherwise, you risk tuning the rest of the pipeline around a bad representation of the source.\n\n*— Louis-François Bouchard, Towards AI Co-founder & Head of Community*\n\n**Our 2024 AI stack did not survive 2026.**\n\nThat is part of why what began as an update to *Building LLMs for Production* became a second book: **AI Engineering for Production**, launching October 20.\n\nThis one focuses less on today’s stack and more on the engineering problems better models alone will not solve: context, retrieval, agents, evaluation, recovery, deployment, and everything required to make capable systems reliable.\n\nYou can join early for book updates, send us topics you want covered, and join the live launch, where we will also share the AI engineering stack we use today.\n\n**Follow the book and get launch details***.*\n\n[Keviv.](https://discord.com/channels/702624558536065165/983037843532308500/1546471713485750312) is building Alineo, an open-source toolkit that gives coding agents isolated environments where they can safely run commands, modify files, install dependencies, and execute code without touching the host system. It can save and resume sandbox state, create snapshots for faster subsequent runs, and fork environments when you want agents working in parallel. It also streams tool calls and execution events back through its TypeScript API, so you can observe what the agent is actually doing inside the sandbox. [Check out the repo](https://github.com/DrejT/alineo) and support a fellow community member. If you have any feedback, [share it in the thread](https://discord.com/channels/702624558536065165/983037843532308500/1546471713485750312)!\n\nTechnically, GPT-6 Astra wins the poll, but not by much. Louie, our CEO, has already moved a lot of his own work to Astra, especially for logic, complex coding, maths, visual analysis, and games. But Fable was his top pick until very recently.\n\nHe has also been using roughly two billion Astra tokens a day over the past week, mostly on client work and internal experiments. That sounds excessive until you look at what the models are actually doing. If an agent can research, implement, test, and keep iterating on essential work tasks, the question becomes less “how many tokens did this use?” and more “was this the right thing to spend them on?”\n\nSo I’m curious: do you have one default model now, or do you already route different kinds of work across several? [Let’s talk in the thread](https://discord.com/channels/702624558536065165/1543076528693190818/1543076528693190818)!\n\nThe Learn AI Together Discord community is flooding with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, [join the collaboration channel](https://discord.gg/rj6m9AF7eC)! Keep an eye on this section, too — we share cool opportunities every week!\n\n1. [Xsnoa](https://discord.com/channels/702624558536065165/784477688551178240/1545341320426561548) is looking for two partners who want to test ideas, build things, launch quickly, and figure out what people will actually pay for in AI. It can be in any niche, such as AI automations, agents, small SaaS products, content systems, etc. If that interests you, [connect with them in the thread](https://discord.com/channels/702624558536065165/784477688551178240/1545341320426561548)!\n\n2. [Bansi_b](https://discord.com/channels/702624558536065165/784477688551178240/1545440943350743173) is working on Python, DSA, ML, and getting into DL/GenAI. He is looking for people who are also learning to maintain accountability, discuss problems, work on projects, etc. If you also want to study with someone, [reach out to him in the thread](https://discord.com/channels/702624558536065165/784477688551178240/1545440943350743173)!\n\n3. [Gcsnotes](https://discord.com/channels/702624558536065165/998978160605540454/1545609460834238484) and his friends are working on a research-based project in the AI memory layer. He is looking for collaborators and AI research enthusiasts; if this sounds like your niche, [get in touch with him in the thread](https://discord.com/channels/702624558536065165/998978160605540454/1545609460834238484)!\n\nMeme shared by [surwren](https://discord.com/channels/702624558536065165/830572933197201459/1545554093957972100)\n\n[Continuous Batching: Why Your GPU Waits for the Slowest Request](https://medium.com/towards-artificial-intelligence/continuous-batching-why-your-gpu-waits-for-the-slowest-request-170dbe2593db?sk=33941d2bd55522af8a3e1dd002dbe8d1) by [Satsawat Natakarnkitkul (Net)](https://medium.com/@net_satsawat?source=post_page---byline--170dbe2593db-----------------------------------------)\n\nStatic batching keeps a fixed group of requests together until every sequence finishes, so shorter requests leave GPU capacity unused while the longest one continues generating. This article quantifies that waste, then shows how continuous batching removes completed sequences and admits waiting requests at every generation step. It also explains why headline throughput multipliers depend heavily on the baseline used for comparison. The useful takeaway is how to judge a serving system using utilization, throughput, latency, and queueing behavior rather than a single speedup number.\n\n1. [A Better Policy Should Not Be Deployed Everywhere](https://pub.towardsai.net/a-better-policy-should-not-be-deployed-everywhere-9d6388da718b?sk=08158a1bfc66c4f8c05e4bee0e19b70d) by [Shenggang Li](https://medium.com/@datalev?source=post_page---byline--9d6388da718b-----------------------------------------)\n\nA new policy can perform better on average and still make outcomes worse for part of the population. This article shows how contextual off-policy evaluation can uncover those cases, especially when the historical data contains too little overlap to estimate performance reliably for a particular group. The author combines cross-fitted doubly robust estimates with confidence bounds to build a gate that switches to the new policy only where the evidence supports it. The result is a practical way to preserve most of a policy’s overall gain without assuming one deployment decision should apply to every context.\n\n2. [Standalone Agent Frameworks vs. Operated Platforms: What a Framework Doesn’t Operate](https://pub.towardsai.net/standalone-agent-frameworks-vs-operated-platforms-what-a-framework-doesnt-operate-c281fc90b59d?sharedUserId=tai-tech) by [MongoDB](https://medium.com/@MongoDB?source=post_page---byline--c281fc90b59d-----------------------------------------)\n\nAn agent framework can define workflows, but it does not automatically solve the operational problems underneath them. This article separates that missing layer into four areas: context selection, observability, scalability, and governance. It argues that model capabilities and standards, such as MCP, will keep changing what belongs inside the framework, while durable state, retrieval, recovery, permissions, and monitoring still need infrastructure beneath interfaces such as Store and Checkpointer. The author uses MongoDB as one implementation of that substrate, then proposes five pass/fail tests around recovery, freshness, and cost to check whether the architecture works beyond a demo.\n\n3. [The Shared-State Problem: When Two Agents Write the Same Memory](https://pub.towardsai.net/the-shared-state-problem-when-two-agents-write-the-same-memory-6854d4ea1ed3?sk=017f98ef780bb74f048b7d60bb97fecf) by [Shrashti Singhal](https://medium.com/@shrashtisinghal?source=post_page---byline--6854d4ea1ed3-----------------------------------------)\n\nOnce multiple agents can update the same memory, familiar database concurrency problems start appearing in agent systems too. This article explains how lost updates, write skew, dirty reads, phantom reads, and stale-context writes arise when agents read and modify shared state concurrently. It then compares approaches ranging from partitioning and single-writer designs to reducers, event logs, optimistic concurrency control, and CRDTs. The important limit is that consistency mechanisms can protect state updates, but they cannot decide which of two contradictory beliefs is semantically correct or prevent the same real-world action from happening twice; those require reconciliation and idempotency at the application layer.\n\n4. [Your Sandbox Shouldn’t Keep Its Install-Time Network Access](https://pub.towardsai.net/your-sandbox-shouldnt-keep-its-install-time-network-access-218096916e60?sharedUserId=tai-tech) by [Divy Yadav](https://yadavdivy296.medium.com/?source=post_page---byline--218096916e60-----------------------------------------)\n\nAn agent may need broad network access to install dependencies, but there is little reason to keep those permissions once it starts executing untrusted code. This article shows how Tensorlake Sandboxes can replace the egress policy of a running sandbox so network access changes with each phase of the task instead of remaining fixed for the sandbox’s lifetime. It explains how allowlists, internet-access settings, failed policy updates, and existing connections behave.\n\nIf you are interested in publishing with Towards AI, [check our guidelines and sign up](https://contribute.towardsai.net/). We will publish your work to our network if it meets our editorial policies and standards.\n\n[LAI #142: My AI Setup in 2026](https://pub.towardsai.net/lai-142-my-ai-setup-in-2026-b5e380f01445) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/lai-142-my-ai-setup-in-2026", "canonical_source": "https://pub.towardsai.net/lai-142-my-ai-setup-in-2026-b5e380f01445?source=rss----98111c9905da---4", "published_at": "2026-09-10 15:01:02+00:00", "updated_at": "2026-09-10 15:17:40.268235+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-products", "developer-tools", "ai-research"], "entities": ["Louis-François Bouchard", "Towards AI", "Claude Code", "Codex", "Obsidian", "AI Engineering for Production", "Building LLMs for Production", "Alineo"], "alternates": {"html": "https://wpnews.pro/news/lai-142-my-ai-setup-in-2026", "markdown": "https://wpnews.pro/news/lai-142-my-ai-setup-in-2026.md", "text": "https://wpnews.pro/news/lai-142-my-ai-setup-in-2026.txt", "jsonld": "https://wpnews.pro/news/lai-142-my-ai-setup-in-2026.jsonld"}}