October 12, 2026
Mind Blow: AI agents that never sleep, a $40 router and a model trained on a MacBook
Always-on agents, cheaper intelligence, local AI hardware and a model trained on a laptop.
A tool called REA claims to let an AI agent reverse engineer any app, down to the binary. Meanwhile a robot arm in Texas was driven live from New York through a web browser.
Here is what happened in AI and DevOps this cycle.
Google launched a single, always-on Gemini agent at its Gemini at Work 2026 keynote. It follows a user across Docs, laptop, phone and the command line with one shared history, and because it runs in the cloud, a multi-hour task keeps going after the laptop goes to sleep. A “dream mode” replays past sessions overnight to learn how the user works. Google also says agents will get their own identity, an audit trail and hard real-time spend caps, so the AI stops borrowing the user’s credentials.
OpenAI put its proactive ChatGPT “dot” on iOS and Android. It has a name, an avatar and a computer of its own. In the demo it started a Codex prototype in the middle of a voice call, then drafted a Slack reply for the user to approve. The assistant now does some of the work, and the user signs off on it. OpenAI also previewed GPT-6 in ChatGPT with Intelligent UI.
A startup is chasing the same idea on the desktop. Aditya Vellanki introduced Sentient OS, on-device AI that he says “understands your whole life overnight” and picks up tasks on the computer with its own cursor. He says it is backed by Y Combinator, Afore Capital and Pioneer Fund. If it works as pitched, the always-on agent moves off the cloud and onto your own machine.
OpenAI is also bringing other companies’ apps into the chat. Plugin extensions for ChatGPT give a product its own sidebar, panel and @mentions, and they can tell ChatGPT exactly what the user has selected in a design editor. It is all built on MCP, the same plumbing behind OpenAI’s own Meetings sidebar.
OpenAI also says code review is no longer the bottleneck. Codex now gets a review inbox for GitHub pull requests, and cloud reviews run automatically and sync their comments back. Each PR gets a private chat that knows the team’s review guidelines, plus an optional local review. GitLab support is next, in preview.
In a two-hour live stream, GitHub’s Wes Bos, Burke Holland and Pierce Boggan had Copilot agents build a fake foldable-phone OS, starting with a home screen and calculator and running several agents in parallel, each in its own worktree. The hosts say the share of agent-written code they keep has jumped from about 50% to more than 90%, and the session cost 2,642 AI credits. Their tip is to get a “rubber duck” second opinion from a model of a different family, and to still read the code.
OpenAI’s DevDay talk Stop Overpaying for Intelligence argued that teams should measure cost per task, not cost per token, because a cheap model that needs human cleanup can cost more. OpenAI says prompt caching makes cached input up to 90% cheaper, and programmatic tool calling cut Clio’s prompt tokens by 38%. One customer went from 24% to 90% cache reuse by keeping the start of the prompt stable. The talk recommends picking models from an accuracy-versus-cost Pareto curve.
GitHub’s Project HydraFusion routes Copilot work across cloud models, and GitHub claims it cuts costs by up to 67%. Copilot will soon auto-discover local models from Ollama and Foundry Local and send simple tasks to them at no AI-credit cost, even offline. GitHub’s suggested split is to plan with Astra and implement locally.
That push to run models locally collides with memory prices. ServeTheHome reviewed the GMKtec EVO-X3, a Ryzen AI Max+ 395 box with up to 128 GB of unified memory that sells for about $3,599–3,799. That is cheap by comparison, because a 128 GB DGX Spark runs about $7,000. It usually draws under 100 W while running LLMs, and the reviewer’s case for local AI is that work under NDA shouldn’t go to the cloud.
Alex Ziskind and Wendell of Level1Techs gave two AI supercomputers a real job: an 8× RTX Pro 6000 rig and an NVIDIA DGX Station running “Turnstone,” a harness that hands work to personas such as engineer, manager and scribe. The DGX Station finished one job in 21 minutes versus 37 on the other machine. The plan is a smart orchestrator directing several cheaper local workers, so one desk can produce the output of a team.
Firecrawl fine-tuned Leia, a 421M-parameter open model, on a MacBook. It took 2,980 search results labeled by Opus and about 98 minutes on an M5 MacBook Pro, peaking at 21 GB of memory. Stock Leia agreed with Claude only 44% of the time, and after fine-tuning it beat Jev head-to-head on this task while running much faster. Firecrawl also built “FireIQ”, a vidIQ clone made in a few hours with agents, which scrapes YouTube pages to markdown and uses an agent-written skill to score titles.
Google’s Gemma 4 open models run on everything from phones to workstation GPUs. One demo, “PenguinAgent”, runs the 26B model offline for field research, and “CodeBuddy” reads a photo of handwritten Python and debugs it with no internet connection. Its sibling EmbeddingGemma 2 maps text, images, video and audio into one space with at most 740M parameters, using 768-dimension vectors that shrink to 128, which is small enough for private RAG on a phone.
Neki Replicator moves Postgres databases tens of terabytes in size at about 0.5–2 TB per hour per shard. It works around a known problem: logical replication holds one snapshot for 20–40 hours and bloats the WAL. Neki instead copies in batches with snapshots of about 15 minutes. The cutover usually takes about five seconds, is capped at 30, and reverts automatically if something goes wrong.
For smaller fires, AWS published a checklist for slow PostgreSQL on RDS and Aurora. It starts with CloudWatch metrics such as DiskQueueDepth, then moves to one-second Enhanced Monitoring and Top SQL, and finally checks dead tuples, autovacuum and lock waits. The rule is to scale up only after tuning. AWS also introduced Lambda Web Functions.
Elastic built an incident-investigator agent that read live logs, metrics and traces and found a latency regression in the frontend proxy, showing every tool call along the way. It is registered in Gemini Enterprise over A2A, and a Cloud Run proxy keeps the API key server-side. Harness showed deployments driven from the IDE: the agent generated a canary-to-Kubernetes pipeline, diagnosed a failed deploy and summarized trends across the last 20 deployments.
Google Cloud showed voice agents that use a real Chrome browser, clicking by x/y coordinates over the DevTools Protocol from a fresh screenshot after every action. Slow actions run in the background so the voice never goes silent. Anything irreversible needs a human to confirm, and the server enforces that with an action token rather than trusting the prompt.
Google Cloud’s guide to scaling agentic coding across a team comes down to three files: AGENTS.md, DESIGN.md and SKILL.md. In a benchmark the talk cites, having an AGENTS.md meant less time, fewer tokens and smaller diffs, and files of about 12 lines worked best. Personal preferences belong at user scope, not project scope. Looker follows the same pattern with five tools for agents. One warning: its CLI acts with the exact permissions of its API key, and enabling a tool on the managed MCP server applies it to the whole instance.
Boundary and HumanLayer argued that coding agents need a memory pipeline. Theirs mines session transcripts for user corrections and the agent’s own fixes, and it caught one BAML fact that tripped up two people across three sessions. They still don’t know how it compares with Cognition’s approach.
Suraj Sharma laid out a six-month plan to become an AI engineer “in the Agentic Era”. Stage one isn’t a framework. It’s token math, context-window limits, context compression and budget allocation, which puts the skill in managing what the model sees. In a post promoting an article, codila claims REA lets an agent rebuild paid app features for free. That claim is unverified and raises obvious licensing questions.
Daily Dose of Data Science listed 11 ways to evaluate an LLM, arguing that no single metric tells you whether a system is good. Leonard Rodman rounded up ten open-source repos that turn AI into a full-time researcher, led by GPT Researcher, and the 0xMarioNawfal account listed six repos for taking agents to production, including a giant awesome-mcp-servers directory. Kushal Gangil pointed to “AI Engineering from Scratch”, a free course he describes as 503 lessons across 20 phases and about 320 hours.
Mori described turning Grok Bot and Opus 5.5 into a 24/7 AI engineering team that coordinates agents, fixes GitHub issues and prepares pull requests overnight. Andriy Burkov noted that LLMs now lean on reinforcement learning, and on a “mid-training” phase of continued pre-training on expert data, to handle code and multi-step reasoning. Ruben Hassid flagged that Google is offering free AI courses with certificates.
Even OpenAI’s community surprised its organizers. The Codex Ambassador program was planned for 15–20 people and now has more than 150 ambassadors in 80+ countries, including one who ran a meetup aboard a moving metro in Kochi. The main requirements to join are curiosity and generosity with your time. At Oracle, business users skip the report request altogether: Codex plans the work, calls MCP services and returns an analysis, a report or even an app.
Jeff Geerling says Qualcomm made a Pi 5 killer, and good luck getting one. The $209 Radxa Dragon Q8B is about twice as fast as a Pi 5 on multi-core work, but its OS images are confusing and uploads cap near 200 Mbps on a 2.5G port. His order took three months to arrive.
Michael Klements found that a $40 travel router, the GL.iNet Mango 2, draws under 5 W over USB-C and can fail over between Wi-Fi, Ethernet, a phone or a cellular modem. From a fake hotel network he reached his home Jellyfin, Immich and Nextcloud over Tailscale. The VPN benchmarks had a clear winner: WireGuard reached about 120 Mbps, while OpenVPN managed only 14–18.
ExplainingComputers made the case for encrypting your data, noting that full-disk encryption also keeps browser session cookies from being stolen. Windows Device Encryption needs a TPM and a Microsoft account to hold the recovery key. On Linux, turn encryption on at install time, and for portable drives, VeraCrypt works on every platform.
On Machine Learning Street Talk, Akarsh Kumar argued that evolution is “anything but random”. By keeping partial solutions, it turns exponential search into roughly linear search. His ASAL work searches a whole space of simulated worlds for the interesting ones, and newer work pits programs against each other in an open-ended arms race.
Julia Turc traced 70 years of rebranding intelligence, starting from an executive order that renames AI “superintelligence” for federal agencies. McCarthy coined “artificial intelligence” in 1955 partly to avoid the word cybernetics, and backpropagation in 1986 revived connectionism, later rebranded as deep learning. Each new name, she argues, has marked a fresh start.
A knitter used Gemini to hide the message “Look what you can do” inside a knitting pattern using carrier matrices. DeepMind’s “Gemini Audio at Night” added TTS with emotion and pacing, plus live translation into about 100 languages. And that robot arm in Texas runs on a Gemini-written script that turns the joint positions streamed from New York into motor commands.
That’s the cycle. Go encrypt a drive, check your cache hit rate, and see you tomorrow.
Also on the radar: these also appeared this cycle, but there was no transcript to report from yet (captions rate-limited by YouTube, no English captions, or no speech), so they aren’t summarized here: Why every company now needs to think and operate like a lab team – Josh Woodward (VP Google Labs) (Lenny’s Podcast); How to run GitHub Copilot agent sessions across any device – GitHub Copilot Day (GitHub); You can monitor EVERYTHING in your homelab with this (NetworkChuck); BigQuery Graph measures explained (Why your AI agents get metrics wrong and how measures fix it) (Google Cloud Tech); Agents in Orbit: Workflows, Not Autopilot — Eduardo Simon & Javier Casal, Flywire (AI Engineer); Reliable, Smart Workflows for Finance, Wealth and Insurance — Karthic Rao, Perpetual (AI Engineer); Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI (AI Engineer); Please… Enough with the Sloppy, One-Shot SaaS Replacements — Mahmoud Abdelwahab, Railway (AI Engineer); Your GPU Neighbor Can Hear Your Model — Branislav Brzak, Stealthium (AI Engineer); Your Agent Is a Lethal Trifecta (And What to Do About It) — Kim Maida, Keycard (AI Engineer); We Let Our Financial Planners Write the Agents — Matthew Ragone, Range (AI Engineer); Building AIE: Open Q&A and Feedback Session with swyx (AI Engineer); Build a Research Agent with Claude Managed Agents (Part 2 of 4) — Anthropic (AI Engineer); Claude Managed Agents Q&A: Evals, CI/CD and Multi-Tenancy (Part 4 of 4) — Anthropic (AI Engineer); Scheduled Agents, Dreaming and Q&A: Claude Managed Agents (Part 3 of 4) — Anthropic (AI Engineer); The Next Trillion Internet Users Still Don’t Have an Identity — Adi Singh, AgentMail (AI Engineer); Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic (AI Engineer); Auth for Agents: Let AI Sign Up for Your Product with auth.md — Michael Grinich, WorkOS (AI Engineer); How Inference Keeps Up with Our Insatiable Need for Tokens — Cerebras & Gimlet Labs (AI Engineer); Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin (No Priors: AI, Machine Learning, Tech, & Startups); AI Is Decompiling Every Game – TheStandup (The PrimeTime); Roundup: Personal agents, AI policy, GPT-6.1 Astra, and Jev decision models (dbt Labs); How Sophos cuts threat response time by 96% with OpenAI Daybreak (OpenAI); Production Monitoring with Codex: Grafana, Kubernetes, & Security (OpenAI); How Oracle Uses ChatGPT Work to Transform Recruitment (OpenAI); Ryzen AI Max+ 495 Hands On! 160GB of VRAM in a Mini PC, GMKtec EVO X5 (ETA PRIME); OpenAI Dots Took Over Our Company (Every); Introducing the Every Agent (Every); VS Code Learn: Spring Boot with Copilot: Build, Debug, MCP & Playwright (Visual Studio Code); BIG Local AI with a MASSIVE NVIDIA DGX Station from Gigabyte (ServeTheHome).