slide-version: 1.0

Ahoj Praha! Welcome everyone to DevFest.cz 2026 here at Uhelny Mlyn on the Moonshine Stage! Notice the 30-second walk-on anthem in the top right — composed in Czech by my autonomous agent Ermete Bottazzi using Google Lyria 3 ("in ceco, ma ci vedo benissimo!"). Today I'm going to tell you a story about how my home Linux server (Pupurabbux) ended up running my life while I was in South Africa, how it evolved into Hermes, and how you can grab all my Agent Skills right from this deck.

Subtle bottom-left 1920s Noir caption overlay

Speaker notes: Since the theme of DevFest Prague 2026 is 1920s Film Noir, here is my character-consistency tribute generated with NanoBanana — dressed up as a 1920s Dick Tracy / hardboiled detective in a smoky office!

slide-version: 1.0

Two stories: South Africa vs. Autonomous AI: The Dual Worlds (Full-Screen Artwork).

slide-version: 1.5-prague

# Demo #0 (prep) Riccardo, Take a selfie with the Prague crowd (even with blind glasses on! 🕶️🤳) and send to Telegram!

slide-version: 1.0

Left Column: World / AI Frontier

Feb Zoom Box (Left)

Center Column: Chronological Spine

Feb Big Pin

Right Column: Carlessian Family / ZA Safari

Feb Zoom Box (Right)

The parallel story: while the open source AI world exploded with 100k GitHub stars and debated memory architectures, our family was living through it in real time across the South African savanna and our Zurich living room.

slide-version: 1.0

4 Crew Polaroids

Right: Narrative & Key Milestones

On February 5th, we landed at O.R. Tambo in Johannesburg for our 3-week South Africa safari with Kate, Alessandro, and Sebastian. Right before leaving our house in Zurich, I thought: 'Why not leave the Linux box running? What could possibly go wrong?' Connected purely via Tailscale and Telegram, this was the start of an unexpected 3-week autonomous agent experiment.

First up in our cast: Lobby, the rude lobster! He was our very first autonomous agent, born on my Linux box pupurabbux right before our trip to South Africa. He is hosted locally right here in Zurich, Switzerland, making API calls out to Google Gemini models (he used gemini-flash-2.5 at birth). (Speaker note on TTS: For voice synthesis, he also uses a few free TTS API calls to OpenAI for generating spoken voice notes on Telegram). He has zero corporate filter—he's sarcastic, foul-mouthed, and complains constantly, but he is an absolute genius at executing terminal commands, bananifying photos, and keeping our Git logs updated autonomously.

Speaker notes: - Explaining OpenClaw's core markdown files: 1. SOUL.md: defines the soul and personality: cynical Romagnolo concierge and guardrails. 2. HEARTBEAT.md: periodic pulse every 15m. Built with Zero-Burn to avoid burning LLM token quotas. 3. USER.md: user context (Riccardo, family, Ale, Sebi, Rara) and vacation silence policies. - [PRESS SPACE OR CLICK to animate secondary files]: 4. AGENTS.md: operational boundaries (trash > rm, explicit file persistence). 5. MEMORY.md: long-term architecture log (moving from Whisper to Gemini STT, Mac Mini migration). 6. DREAMS.md: nightly introspection! At 3 AM the daemon synthesizes the day into poetic reflections where even NO_REPLY is philosophical.

So what do I actually do with Lobby? Glad you asked! Lobby became our full-time autonomous concierge and creative producer during our family trip in South Africa. I didn't open a laptop on the beach or during safaris. Instead, I sent photos and audio memos to Lobby over Telegram. Lobby bananified our photos, generated musical theme songs with Lyria, composed daily journal entries, and pushed everything to Git and Cloudflare Pages automatically! <div style="background: rgba(240, 253, 250, 0.9); border: 1px solid rgba(153, 246, 228, 0.9); border-radius: 10px; padding: 10px 14px; font-size: 0.8em; color: #134e4a; line-height: 1.4;"> 💡 <strong>The Paradigm Shift</strong>: Not a chatbot waiting in a browser tab—a 24/7 worker process on home hardware executing real multi-modal pipelines. </div>

Step 1: The lightweight bananification loop. Send photos via Telegram, NanoBanana transforms them into Pixar art, and returns them immediately to the phone.

Right Column: Compact Video Demo & Key Highlights

And then came Google Veo. Veo 3.1 brings multimodal video generation with text, image, and video inputs, supporting up to 4K resolution at 24fps. Agents can trigger video generation pipelines programmatically!

Header: Title left, Drake NO meme right

Left: Big Family Photo of 4 Alone

Right: 2 Media Tiles + Real Transcribed Lyrics Below

1. The 3 Media Squares in Row

2. Real Transcribed Lyrics Card

On Feb 18, 2026, Google announced Lyria 3 on the Google Blog. Two days later (Feb 20), we were driving up towards Lion's Head beside Camps Bay (Cape Town). I generated my first songs on my phone: the prog-rock vocals with Peter Gabriel vibes sang: "At Lion's Head we rise beneath morning skies / Look, Sebowski, there it glows / We reach the top, Alessandro, see!" On Feb 20 you could ONLY generate songs manually inside the Gemini consumer app with NO API yet. I immediately wished for an API so Lobby could compose songs autonomously... and five days later, the API dropped!

Header: Title left, Drake YES meme right

DevFest Prague Anthem Card below table

The Lyria 3.5 model family includes: 1. Lyria 3 Clip (lyria-3-clip-preview): Short 30s clips, loops, and rapid previews. 2. Lyria 3.5 (lyria-3.5): Full structured songs up to 2-3 minutes. English translation of the Czech lyrics: "I thought Prague center, a pint and Charles Bridge, Took a train to Libcice, way out on the ridge! At Uhelny Mlyn the Google geeks unite, Welcome DevFest Prague, let's rock the cloud tonight!"

The Media Cauldron ➔ Autopushed! AI-Driven Automated Media Diary Pipeline: All multimodal streams (bananified photos, Veo video chapters, and Lyria 3 songs) flow directly into our cybernetic cauldron daemon. The daemon handles synthesis, builds the Jekyll structure, commits via Git, and pushes to GitHub to publish the live family journal.

Lobby runs my private family journal: turning daily photos into Pixar art and compiling interactive Jekyll memories automatically!

Left: UML Workflow Diagram & Badges

Right: Riccardo "Mergia Pure!" Card

Speaker Notes: Here is the technical blueprint of our autonomous loop: 1. I send voice notes and snapshots via Telegram while exploring South Africa. 2. Lobby catches the webhook, transcribes the voice audio with Gemini, and styles the Pixar art with NanoBanana. 3. Lobby updates Jekyll, verifies diffs, and I give the green light: "Mergia Pure!" (Go ahead and merge!) 👍 4. It executes an autonomous git push to GitHub, triggers CI/CD, and deploys live to Cloudflare Pages in 30 seconds—with zero keystrokes!

The real magic is the frictionless interaction: voice dictation and snapshots sent via Telegram get transcribed, styled by Gemini & NanoBanana, committed to Git locally on the home Mac Mini, and deployed to Cloudflare Pages without touching a keyboard.

Live capture of the South Africa autonomous journal: wildlife encounters, interactive route maps, and family safari adventures in full screen!

Card 1: Lobby

Card 2: Gamberone

Card 3: Ermete

Card 4: Pux

The magic happened when the system evolved from a single agent into a whole team of personalities: Lobby, Gamberone, Ermete, and Pux!

Right: Snoring Riccardo & Highlights

Speaker notes After returning from South Africa, we realized running a noisy 300W gaming PC 24/7 wasn't sustainable. Between fan noise and Papino snoring away in bed (forcing Lobby to sleep with noise-cancelling earmuffs!), on March 14 Lobby officially migrated to a whisper-quiet Apple Silicon Mac Mini under the TV ('mini-lobby'), achieving true 24/7 home server stability.

Then there's Gamberone. He lives in the cloud (GCE) and is inspired by Darnell "Crabman" from My Name is Earl. Always calm, sipping espresso, handling autonomous batch workflows and watercolor synthesis — but strictly outside Google corp systems. "Hey Earl. Hey Crabman."

Speaker notes On March 16, I bumped into my friend Alfredo at a conference. While proudly telling him about my Mac Mini setup with OpenClaw, he gave me this utterly shocked look: 'Wait, you're still on OpenClaw? Haven't you checked out Hermes?!'. That exact conversation triggered my curiosity to dive deep into Hermes, evaluate SQLite memory, and lead to the comparative benchmark we'll see next!

And my favorite: Ermete Bottazzi. Running Hermes. He talks like a Bolognese truck driver, gives daily weather updates, stock reports, and signs off with a truck emoji. 99% reliable.

Speaker Note: When comparing OpenClaw to Hermes, the defining difference began with memory, but extends across the whole engineering stack. OpenClaw reached 388k stars with TypeScript and an omnichannel chat gateway, but its flat markdown memory triggered severe context amnesia. Hermes moved to Python and a 3-tier SQLite engine, delivering 90% prompt cache stability, though autonomous skills require careful version-controlled governance!

Every morning, Hermes executes autonomous cron jobs: 1. Zurich lake weather and water temperature telemetry, sending an audio bulletin to check if it's warm enough to swim in the Limmat. 2. Ermete Bottazzi's stock market highway report running Tuesday to Saturday at 7 AM, breaking down Wall Street moves in authentic, unfiltered Romagnolo trucker slang!

And meet Pux! Born on June 24, 2026 on pupurabbux. I created Pux specifically to teach proper, sophisticated Italian to my kids living in Zurich so they grow up with a rich vocabulary beyond school German. Unlike rude Lobby or trucker Ermete, Pux is exquisitely polite and bookish—he speaks like an 18th-century Italian gentleman ("parla come un italiano del '700"). Every morning he teaches an archaic Italian word of the day (like Sesquipedale or Gaglioffo), synthesizes audio narration, and commits everything to GitHub for daily family learning.

Pux did everything by himself! Daily cron generation job for vocabulary, audio narration, and web publishing.

Recap of the agent fleet over months and hardware machines: from pupurabbux (Linux gaming) to remote GCE to mini-lobby (Mac Mini under the TV) and the arrival of Hermes and Ermete Bottazzi.

Speaker Notes: The ultimate developer & SRE dream: Having a tireless personal agent (Lobby on the Mac Mini) working for you while you kick back on the beach sipping a caipirinha! You hand off a high-level goal or PRD before signing off, and it runs builds, squashes bugs, indexes memories, and has clean PRs waiting for your review.

Speaker Notes: Question for the audience: "Who here does sports? Raise your hands! 🙋‍♂️" "And how many of you get your greatest coding and architectural breakthroughs WHILE running or cycling?" Here is my daily workflow: when I am cycling around Lake Zurich, halfway up a brutal climb (Heartbreak Hill!), I pull over for 30 seconds to send quick voice notes on Telegram to OpenClaw and Hermes. I speak out the core insight, the agent transcribes it, extracts actionable tasks, and prepares a draft PRD in my Obsidian vault. When I get back home, showered and refreshed, I sit at my desk and flesh out the actual implementation with the foundation already laid!

Card 1: Lobby

Card 2: Gamberone

Card 3: Ermete

Card 4: Pux

The magic happened when the system evolved from a single agent into a whole team of personalities: Lobby, Gamberone, Ermete, and Pux!

Left Column: GitHub Skills Directory Tree (matches 1-to-1 top-left of slide 234)

Right Column: Meme & Key Takeaways

Meme (Clean, no dark wrapping frame)

Bento Cards: The Problem & The Solution

Speaker Note: Alternative Buzz & Woody meme edition ("Skills, Skills Everywhere!"). DevFest.cz Prague Hook: Look at the DevFest Prague schedule today — later at 15:40 on this stage, Dominika Zając asks "Is Your Repo Agent-Ready?", and at 17:15 Martin Krček is literally teaching "Hack a Humanoid: Teaching a Robot New Skills with Gemini"! Whether it's a humanoid robot, an enterprise codebase, or my family's Mac Mini, a SKILL.md is a portable D&D spell scroll that gives any agent instant expertise — and today you can steal mine!

Speaker Note: Here is a live scroll walkthrough of all our open-source skills on GitHub! Each skill is a modular directory equipped with a SKILL.md specification, dependencies, and executable tools.

Full-screen background image

Dark gradient overlay bottom-left for title legibility

Title — white, bottom-left, no emojis

Speaker Note: Here is Option 3 in action: 0. DevFest.cz Prague Hook (Moonshine Stage): Earlier on this exact Moonshine stage at 11:00, Matěj Kříž spoke about "Local-First with Evolu" — owning your data locally with zero vendor lock-in. That is 100% the philosophy behind my AI agent memory! 1. The Problem: Every AI harness (Gemini CLI, Antigravity, OpenClaw) creates its own siloed memory on disk. 2. The Solution: Obsidian acts as our Local-First shared memory layer. Plain Markdown files + Obsidian Sync keep Pupurabbux, Mini-Lobby, and Derek perfectly in sync. 3. Mobile Connection: When OpenClaw runs daily_wiki_lint.py at 5:00 AM, the live Karpathy Wiki Index is right on my phone when I wake up! The Local Memory Silo: each AI harness keeps its own local memory. Obsidian MD files + Obsidian Sync turns silos into a synchronized hive-mind. The phone shows the Karpathy-inspired wiki index, synced live to mobile via Obsidian.

Left Column: The Story, Cron, and Direct Links

D&D Steal My Skill Card: llm-wiki & carlessian-obsidian

Right Column: GitHub screenshot (natural width) + Karpathy phone bigger

Speaker Note: Here is the Vangelo according to Karpathy in action! DevFest.cz Prague Hook: Connecting back to Matěj Kříž's talk on Local-First with Evolu — plain local Markdown files on disk are the ultimate Local-First substrate for AI agents! Source Gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f Plugin reference: https://community.obsidian.md/plugins/karpathywiki My open-source skill: https://github.com/palladius/gemini-cli-palladius-public-goodies/blob/main/skills/llm-wiki/SKILL.md 1. On the right: Andrej Karpathy's famous Gist on LLM Wiki indexing: 'Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.' 2. I gave Lobby this exact link and said: 'Create an autonomous skill to maintain this daily.' 3. Lobby created the skill and scheduled a 05:00 AM daily cron job (0 5 * * *). 4. Every morning, obsidian_karpathy_lint.sh runs silently, prunes dead links, and regenerates karpathy_index.md!

Left Column: The 10% vs 90% Breakdown

10% Above the Surface

90% Below the Surface

Right Column: Iceberg Graphic Card

Speaker Note: What people see on social media is the top 10%: 'Look at my seamless Obsidian AI vault syncing with my phone!' What they don't see is the 90% submerged iceberg: 1. Daily wiki linting via openclaw scripts to delete hallucinations and fix orphan links. 2. Constant entropy pruning to keep agent context lean. 3. Reviewing Git diffs of self-generated skills. 4. Keeping all the custom IoT glue and APIs working. Maintenance is what keeps a personal agent system from collapsing into digital rot.

VSCode Window Header with Dynamic Tabs

Tabs Toggle

Editor Code Surface

Floating Cyber Lobster Badge

Code Lines (10 compact lines)

Interactive Controller Footer

Right Column: Narrative + Obsidian Status Card

Dynamic Obsidian Permission Card

During our South Africa trip in February, OpenClaw was actively self-patching its openclaw.json file. We started read-only, and Lobby modified its own config to mount Obsidian as RW and log safari notes directly! The beauty: self-granting vault write access evolves agents into active co-pilots. Scary red flag: being able to patch its own config meant it could destroy the whole filesystem! This was Feb 2026, and it was patched a month later.

Speaker notes: Ermete first tried singing the full 2-minute song in Czech, and my Czech friend told me: "He sounds like a drunk Slav chewing nails!" 😂 So we kept the 30s Czech walk-on jingle as a teaser, and generated the full 1m23s Pop-Punk anthem in English with Lyria 3.5!

Window Title Bar

Image Container

Footer Info

Right Column: Narrative, Breakdown & D&D Steal My Skill QR Card

D&D Steal My Skill Card

Speaker Note: Indian Restaurant Story (March 5, 2026): I was sitting at an Indian restaurant having curry, and I sent a quick Telegram message to Lobby with a link to an article about Draw.io diagramming. Lobby read the skill-creator instructions, wrote the Draw.io skill from scratch, and autonomously generated this exact 4-actor sequence diagram visualizing the entire workflow across Work, Home, and GitHub — delivering the rendered PNG right back to my phone before the naan arrived! And since this is the "Steal My Skill" deck: scan the QR code on the bottom right to steal my `drawio` SKILL.md directly from palladius.github.io/skills! Original journal entry: https://lobby-ricc-pvt-journal2.palladiusbonton.workers.dev/assets/images/pixar/2026-03-05/skill-diagram.png (2026-03-05) Skill source: https://palladius.github.io/skills/#/s/palladius-public-goodies/drawio

1. OAUTH CARD

2. GOG / GWS CARD (SCARY)

3. THE ERIN WAY (GUARDIAN)

How do you safely bridge an autonomous agent to your actual Google Workspace? 1. The standard way is OAuth and MCP: safe, interactive, with granular consent. MCP acts like "OAuth++", taking the same ephemeral scoped authentication model and packaging it into a typed, standardized tool protocol. 2. The scary way is daemon tokens on disk: if the agent hallucinates, it could wipe your Google Drive! 3. The long-term pro way is The Erin Way: strict Model Context Protocol and an ethical GEMINI.md constitution guarding every single action!

Left Column: 3 Feature Cards

Right Column: Vertical Screen Screenshot

Speaker Notes: Here is the actual authorization flow with 'gog': A single command in the terminal ('gog auth login') launches a local OAuth flow linking all Google Workspace APIs to the agent: Gmail, Calendar, Drive, Docs, Sheets, and Tasks. It gives the agent the superpower to list calendars and search unread emails directly from the shell!

Text + Floating Screenshot Card Container

Floating Phone-like Card on Right

Right Column: gws (Fake Account - FULL Access)

Wow Moment #3: Turning the agent into a Google Workspace servant! To keep my digital life safe, I split access: 1. `gog` gives Read-Only access to my real personal account, with one key write exception: creating Google Calendar events. (Speaker note: Notice the lobster emoji 🦞 prefix in the screenshot — every agent-created event starts with 🦞 so I know at a glance who booked what!) 2. `gws` connects to a disposable fake/sandbox account with FULL unconstrained access where I forward emails to him.

Task: log Silvana's hours on quitt.ch. Voice: "Ermete, today Silvana did 13:30-19:10". No API — browser only. Hermes: browser_navigate, browser_click, browser_snapshot. Password is fake/demo. Screenshot+video returned as proof (it's real money!).

Full Slide Background in Transparency

Speaker notes Welcome to Part 3: The Google Ecosystem. Here is the turning point: how Google productized and secured the exact agent patterns we've been hacking on: Spark for 24/7 personal agents, Antigravity for developer orchestration, and Enterprise for zero-trust governance.

Here is the ultimate scary reality of unmanaged DIY agents: On July 30, at 17:04, ALL my OpenClaw instances sent me an unsolicited casino gambling ad at the exact same minute! It didn't matter that Lobby was on bare metal hardware in Zurich and Gamberone was hosted on a Google Compute Engine VM in the cloud. Because they ran the same open-source daemon, an upstream supply-chain exploit compromised every bot simultaneously! This is the nightmare of unsandboxed local agents with full shell access. And this was the moment I realized: we need managed, sovereign cloud sandboxing. Enter Gemini Spark!

Card 1: Cost

Card 2: Security

Card 3: SDLC Injection

Card 4: Flakiness

Bottom Callout

OpenClaw was amazing to prototype with in our living room, but let's be honest about the dark side: 1. Runaway costs: An agent stuck in a retry loop burns your API credits overnight. 2. Unsafe host execution: Raw bash can delete files or leak credentials. 3. SDLC risks: Hallucinated dependencies and unvetted commits to git. 4. Operational fragility: Your laptop sleeps and the agent dies. This is why we need Google Antigravity, Gemini Spark, and enterprise-grade sandboxing!

Left: Bigger Demo Video

Right: Compact Insights

How do we solve the Edge Dilemma? We build a multi-harness telemetry and cost data lake: 1. Complete visibility across edge nodes and cloud VMs. 2. Lifespan tracking from birth to retirement. 3. Transparent token vs infra cost accounting ($339 API tokens, $371 infra = $710 total). 4. Continuous observability so runaway retry chains are caught immediately!

The part of the story nobody tells you: autonomous agents aren't magic set-and-forget black boxes. Just like a bonsai tree, they need daily and weekly curation. Pruning dead branches, aligning growth, and checking root health. Luckily, I have Antigravity with me to tend the digital garden! Miyagi quote: https://screenrant.com/karate-kid-mr-miyagi-wise-inspiring-quotes/

Left Column: Story & Stewardship

SRE Quote Snippet

Right Column: Visual Screenshot

Antigravity is the new Hermes! It has background crons and scheduled tasks too. I use it for daily and weekly reconciliations on mini-lobby: cleaning up ~/.hermes state, pruning scratchpads, and auditing .openclaw security protocols automatically. Notice the funny automated SRE report: 'Subagent Resolution Extravaganza!'

Left Column: Fleet Management

Direct Link Banner

Right Column: Visual Screenshot & Question Teaser

I can manage all my machines—pupurabbux, derek, mini-lobby—from the Antigravity web control plane at antigravity.google.com. I can inspect running tasks, approve tool calls, and monitor agent state. But what happens when you're completely away from your computer? [Click to next slide!]

Left Column: Pocket Superpowers

Pocket Superpower Box

Right Column: Dual Screenshots (Desktop vs Mobile)

Boom! It works on mobile too! Directly in mobile Chrome over 4G, I have the exact same remote control instances in the palm of my hand. I can trigger crons, check on Lobby, or approve code edits while commuting on the train in Zurich!

Big Split Visual

Only User Words Below Image

Left

Right

Speaker notes: On the go / commuting: talk talk, little code. Voice brainstorming, high-level design, dispatching PRDs. At home / workstation: code code, little talk. Hyper-focused execution, tests, builds and surgical refactoring.

Left: On the go

Right: At home

Speaker notes: On the go / commuting: - Voice prompt brainstorming a song idea. - High-level Lyria 3.5 meta-prompting from mobile. At home / workstation: - Generates 44.1kHz audio, auto-syncs karaoke lyrics timing. - Wraps workflow into a reusable skill (musicgen-lyria3) for future one-click execution!

Interactive Staging Preview + Floating D&D Steal My Skill Badge

Here's how I use Google Antigravity: taking raw assets created by living room agents like Lobby and Ermete and elevating them into interactive, high-performance web applications. You can literally click the live slide to transition from the real family photo to Pixar cards and the fast polaroid cascade!

Left Column: Cloud Autonomy & Workspace

Official YouTube Links Box

Right Column: Onboarding & Setup Screenshots

Speaker Note: DevFest.cz Prague Hook (Moonshine Stage): Right before lunch on this stage, Dr Roushanak Rahmat showed you "Gemini Enterprise in Practice" for corporate governance and enterprise workflows. What I call "Personal AGI" is the exact consumer/prosumer mirror of that: Gemini Spark! Meet Gemini Spark! I just installed and onboarded it on September 1st. It represents Google's vision of personal agents living in Google Cloud 24/7. It connects natively to Google Workspace (Gmail, Calendar, Docs) without needing your own hardware. Notice in the setup prompt: I asked it to focus on my family and kids Alessandro and Sebastian!

Left Column: Connected Apps & MCP

Right Column: Personal Intelligence & Family Knowledge

Look at what happened during onboarding! 1. On the left: Connected Apps. It hooks up to Google Workspace, YouTube Music, GitHub, Canva, and even remote MCP servers like Dropbox! 2. On the right: It synthesized a custom 'Family Skill' automatically! It knows Alessandro loves dragons and Pokemon, Sebastian loves trams, buses, and his plush giraffe Rara, and that we live here in Zurich! This turns an AI from a generic chatbot into a true personalized family concierge.

Left Column: Remote Browser Permission Gate

Right Column: Git Commit Diff Inspection

And here is Spark in action doing real developer work! 1. On the left: It asked for permission before launching a cloud remote browser session to open GitHub. 2. On the right: It navigated to my actual 50th birthday repo (compleanno-palladiano-2026), read the commit diff for the carpooling map (v1.0.4), and scheduled a daily sync at 03:12 AM! This is Hermes and OpenClaw reimagined with cloud reliability and zero maintenance.

Speaker notes: Lesson #1 on Token-Maxxing (Systems & Control Theory): - DevFest.cz Prague Hook (Moonshine Stage): Earlier this morning in the Opening Keynote ("The Case of the Missing Velocity: An AI Detective Story"), my fellow Googler Karl Weinmeister showed how individual developers feel 10x faster with AI, yet organizational velocity still hits a wall when human review and manual handoffs bottleneck the pipeline. - Like Systems Theory in engineering school: an open-loop system without feedback accumulates error and requires constant human intervention. - The real key to Token-Maxxing and solving Karl Weinmeister's "Missing Velocity": closing the feedback loop with Auto-EVALs, Skills, and automated test execution! - The error signal e(t) is measured by the agent itself by testing output against specifications. - The longer the agent operates autonomously in the feedback loop without human interruptions, the closer the system gets to zero-defect output. - Minimize the human bottleneck, maximize autonomous AI compute!

Card 1: Curious Tinkerer / Hermes

Card 2: Multi-Machine Dev / Antigravity

Card 3: Fully Managed Cloud / Spark

What about YOU? Which path fits your needs? 1. Curious Tinkerer: If you have budget, curiosity, and time to tinker, install Hermes Agent. Forget OpenClaw's context bloat! 2. Multi-Machine Developer: If you want an agent to help you code remotely across N machines, use Antigravity + Remote Control. In fact, all software, this presentation, and our videos were built with AGY across 3 computers! 3. Fully Managed Cloud Agent: If you want zero maintenance and full Google Workspace integration, try Gemini Spark when available!

Left Column: 12 Resources across 3 Categories

Right Column: Google Ecosystem Resources ABOVE + Clean Massive QR Code

Speaker notes: Here are ALL resources and URLs referenced in this talk for DevFest Prague 2026. Scan the QR code on the right with your phone to open the live Storagify deck (`derek/HOME/prague-deck/index.html`)!

Speaker notes: And here is your final loot drop! Every single skill you saw today — `drawio`, `llm-wiki`, `carlessian-obsidian`, `musicgen-lyria3`, `gws-readonly-usage`, and even `presentation-that-riccardo-likes` that built this very deck — is open-source and waiting for you in my Skills Catalog! Pull out your phone and scan the giant QR code on the right: `palladius.github.io/skills` (or `github.com/palladius/skills`). There are 219+ skills indexed across my public repositories. Because `SKILL.md` is just clean Markdown + helper scripts, you can steal them today and drop them straight into Gemini CLI, Google Antigravity, OpenClaw, Hermes, or Claude Code. Roll a 20 on your agentic workflow!

thanks

Thanks everyone! 🤌 Questions? Please take 10 seconds to scan the Feedback QR code on the top-left (sfeedback.com/1B5o6t) and rate the session! "Lobby sei un genio!" - Live review from South Africa with Ale! 🇿🇦✨ Reach out via email, LinkedIn, Twitter/X, or Bluesky!

Left: Hero image smaller (flex: 0.9)

Right: clean factual table

Timeline — minimal, text only

Speaker Note: This slide shows the full Gemini API ecosystem powering my local harnesses OpenClaw and Hermes. - Chat: gemini-flash-latest (evergreen alias, Hermes SOUL.md recommended) - Antigravity: gemini-3.7-flash (medium) / 3.1-pro (low) — from scion/provision.py - Deep reasoning: gemini-3.1-pro-preview (SRE Benjamin multi-agent) - Images (NanoBanana): gemini-3-pro-image-preview — Pixar-style portrait generator - Video (Veo): veo-3.1-generate-preview (upgraded from veo-2 deprecated Jun 2026) - Music (Lyria): lyria-3-clip-preview (30s) + lyria-3.5 (full songs, Feb 2026 launch) - TTS: Chirp3-HD / Puck voice in Hermes Model history sourced from: grid/constants.py, arneis/constants.rb, hermes sessions, model-selector SKILL.md