Carlos Sanchez

29 July

Packed room for our #KubeCon Japan. #ArgoRollouts with AI Agent orchestration to autofix production issues while safely doing canaries with @kevindubois.com ありがとうございます kubecon-cloudnativecon-japan-2026.sessionize.com/session/1145...

Looks legit #addisababa

Docker Maven images are now ready for 3.10 on Eclipse Temurin, for Alpine and Ubuntu Resolute. Resolute is also available for all Temurin variants. Example: maven:3.10.0-eclipse-temurin-27-resolute

harnesstax.github.ioHarnessTax: How Much Does the Harness Matter for Coding Agents?What does a coding-agent harness actually add, and at what cost? It turns out your Claude models may not need Claude Code… We evaluate 21 model–harness pairs spanning seven models and three harnesses—...

18 Sept

🤔 "Harness choice has little effect on task success rate, but can significantly affect the cost" "A simple harness can be competitive" "Models may perform better with other harnesses than with their own. So it turns out that your Claude models may not need Claude Code…" harnesstax.github.io

Lots of interesting takes here. One thing I miss: the perf factory should also address business outcomes (we are doing that with usage data analysis)

Gergely Orosz@gergely.pragmaticengineer.com

Here's what OpenAI's agentic software factory looks like, today. It's surely not token efficient, but eg agents "babysitting" releases to prod is something new, and an interesting approach. Details: newsletter.pragmaticengineer.com/p/openai-sof...

> Death of the IDE & pull requests. PRs and code reviews are being rethought > Agents handhold both the code changes and the changes behind feature flags. Builds its own (feature flags) monitoring dashboard to use

> We are building an agent that “handholds” a change all the way to production Low-risk changes follow an easier path; areas of the codebase can opt in to an agent that will auto-approve low risk PRs, removing human acceptance as a bottleneck and improving velocity.

youtube.comAgentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, AdobeCarlos Sanchez types a request for a coffee machine he can use while camping, and the page assembles itself in under two seconds. Not a search result. A page, with camping appropriate machines,…

Tested Jev to see what the buzz is all about and this is what I found for my generative websites case 🧵 TL;DR: it's not a chatbot. It's a "choice machine." And for the right job, it's stupidly fast, cheap, and accurate No free text generation, ever.

My generative websites demo bsky.app/profile/csan...

Carlos Sanchez@csanchez.org

My talk "Agentic Sites: Building Hyper Personalized Websites" from AI Engineer SF is now up I show a demo of site content being autogenerated live in under ~1.5 seconds, opening the door to a extremely personalized web experience, while keeping the content on brand www.youtube.com/watch?v=jebp...

typesafeai.bsky.socialDiogo Almeida (@typesafeai.bsky.social)Neolab

Jev -> typesafe.ai typesafeai.bsky.social

Regular LLMs are like asking someone to write an essay. Jev is handing them a multiple-choice bubble sheet. If your task IS actually a bounded decision (classify, rank, select, score) you don't need an essay. You need a bubble sheet.

Not a chatbot replacement, a decision-engine replacement. Cheaper, faster, and often just as accurate, for that slice of the pipeline.

Intent classification: also 15/15 (100%) tweaking the process. Instead of generating the response text make Jev answer if product X was mentioned (yes/no) The one real limit I hit: fully open-ended extraction (arbitrary "use cases"/"features" with no fixed list). Jev genuinely can't do that

💰 Cost check: ran it against ~28 other providers (Bedrock Claude/Nova/Llama, Cloudflare Workers AI, self-hosted models). Jev landed in the CHEAPEST tier on par with micro-models, while being the fastest 100% on block selection. ~80-400x cheaper than Claude Sonnet/Opus for equivalent accuracy.

My site has a generative pipeline with 2 stages: 1️⃣ Classify user intent + extract entities 2️⃣ Pick which UI blocks to render (hero, comparison table, testimonials, etc.) Both are "decision" tasks, not "writing" tasks. Perfect test case for Jev

Block-selection results: 15/15 (100%) ✅ I turned "which blocks should render?" into 21 parallel yes/no questions (one per block type), answered in a SINGLE API call, then combined with plain code logic. Matched Claude Sonnet 4's accuracy at ~4.4x the speed.

It only answers 3 question types: 🅰️ Choice (pick 1 of N) ✅ Noul (yes/no, 0-1 confidence) 🔢 Score (numeric rating)

I also show what a experience could be with the integration of multi modality: talking to your voice home assistant, getting results on your Google TV. No more science fiction. This was built on Adobe AEM, Cerebras fast inference and Google Gemma 4 model

labs.googleDiscoTake the web for a fresh spin

25 Aug

Giving a try to Google Disco Browser. Interesting concept, site personalization for the individual user is the future. Similar concepts to what we are building with "Audience of One" aka Generative Websites labs.google/disco

5 Aug

Big responsibilities in this flight emergency exit "Hold other passengers back from the exit while crew operates to open"

30 July

Thanks for the feedback too! ありがとうございます #kubecon

Carlos Sanchez@csanchez.org

22 July

こんにちは Next week I'll be at #KubeCon Japan with @kevindubois.com talking about using #ArgoRollouts with AI Agent orchestration to autofix production issues while safely doing canaries. A practical AI actionable use case with low risk! kubecon-cloudnativecon-japan-2026.sessionize.com/session/1145...

Back from speaking at @aidotengineer.bsky.social in SF last week and #googleioconnect Berlin just before. Lots of energy and cool (AI) stuff * The self-improving codebase is real * The agentic org chart * Agentic web standards * Agent sandboxing * Open models, training, fine-tuning #AIEngineer

Models, Training and datacenters: plenty of companies selling inference, model routing, training, post-training, fine-tuning,... on open models

Great talks on the hallway track with @boredabdel.bsky.social @cloudtaquero.bsky.social @markvillacampa.com @rustam.no @ricmac.org @charles-irl.bsky.social @peopleforrester.bsky.social @doble.io @liad.bsky.social and many more

The agentic web is getting standardized. WebMCP (Chrome): sites expose in-browser tools directly to agents, no DOM scraping. MCP Apps: servers push interactive UI back into agent chats. ARD/ai-catalog.json: agents discover tools & agents at runtime. auth.md agents register users without signup forms

Agent sandboxing is the new infra fight: Gemini Enterprise Agent Sandbox, AWS Lambda Sandboxes, k8s agent-sandbox and substrate, Nvidia OpenShell, Modal, E2B,... Workload isolation + fast snapshot/restore

The self-improving codebase, aka Dark Factory, is real, but old rules still apply. Good CI/CD, e2e tests & canaries directly speed up the agentic loop — skip them and humans stay the bottleneck. ie. Uber already write, review, test & merge PRs with barely any human touch

The agentic org chart. Forget managing multiple tabs — the new model is for one human to interact with one "AI manager" that manages multiple agents. Cloud dev environments let those sessions run nonstop at scale, no need to keep your laptop open.

latent.spaceThe website of the future may assemble itself for every visitorAdobe is experimenting with “agentic sites” that generate pages around an individual user’s intent. At AIEWF, we talked to Carlos Sanchez about the Web's future.

7 July

I had a chat at AI Engineer with Richard about the future of the web: agentic hyper-personalized sites, WebMCP, MCP Apps, A2A,... #AIEngineer www.latent.space/p/the-websit...

6 July

3 days, 2200+km into Oregon. Beautiful

4 July

My fellow Americans, we have to talk about Dutch Bros. Why? 🤌

presentations.csanchez.orgpresentations.csanchez.org

2 July

Is the future of the web chat interfaces? I don't think so presentations.csanchez.org/2026-06_ai_e... #AIEngineer

presentations.csanchez.orgAgentic Sites: Building Hyper Personalized WebsitesAgentic Sites: Building Hyper Personalized Websites

2 July

My slides from #AIEngineer Agentic Sites: Building Hyper Personalized Websites presentations.csanchez.org/2026-06_ai_e...

1 July

Token Billionaires @aidotengineer.bsky.social

1 July

Building is easier, generating value is still hard

1 July

Fable is back* *Later today @aidotengineer.bsky.social

30 June

The future is not 20 terminals, it is better loops. Peter Steinberger @aidotengineer.bsky.social #AIEngineer

x.comSam Hogan 🇺🇸 (@samhogan) on XWant to try GLM 5.2 in production but worried how it might change your product? Don’t worry, we got you: 1. Install Inference Gateway (https://t.co/LZqoPEz3Sr) 2. Keep sending traffic to your curren...

30 June

Very good to see good engineering practices now applied to AI, traffic cloning #progressivedelivery x.com/samhogan/sta...

@monkchips.com fyi

Token Billionaires @aidotengineer.bsky.social

29 June

25h to get from Madrid to SF but ready for @aidotengineer.bsky.social, send ☕️ #FML

25 June

At Google I/O in Berlin this week and AI Engineer World’s Fair in SF next week I will be presenting Adobe frontier work on using fast inference to individualize websites as you visit them - Audience of One of1.live #googleio #aiengineer

blog.csanchez.orgTesting Kubernetes sandboxing technologiesThere are several flavours of “give a workload its own isolated runtime on Kubernetes” floating around right now. We wanted a head-to-head comparison — not a slide-deck comparison, but …

23 June

Testing and comparing #Kubernetes sandboxing technologies for AI Agents: agent-sandbox, openshell, substrate, kars blog.csanchez.org/2026/06/23/t...

22 June

Life is what happens between tokens budget exhaustion until the next cycle

github.comGitHub - carlossg/playwright-k8s-sandbox: Running playwright in a sandbox in k8sRunning playwright in a sandbox in k8s. Contribute to carlossg/playwright-k8s-sandbox development by creating an account on GitHub.

17 June

If you’re evaluating different sandbox solutions for #kubernetes, check out playwright-k8s-sandbox to test and experiment with running Playwright in a sandboxed environment. Comparing agent-sandbox, openshell, substrate, and kars

youtube.comThe case of ArgoCD and CDEvents integration | Argo Unpacked Ep. #20YouTube video by Argo Unpacked

8 June

Going live in 20 min 🚀

Argo Unpacked@argounpacked.bsky.social

🎙️ Argo Unpacked – Episode #20 is happening today! We’re joined by Ben, Kostis, and Luke to dive into the case for Argo CD and CDEvents integration and what it means for the GitOps ecosystem. 👉 www.youtube.com/watch?v=p-gB...

6 June

I've burnt 80% of my AI budget for the month in one day building a PoC with #kubernetes agent sandbox and agent substrate. Totally worth it, will post details. And I can always get more budget 😀