Blog
Web development, AI, experiments and whatever comes up.
AI Agents 2026: Open Source Governance & Lead Intelligence
This week the agent ecosystem split. Microsoft and OpenAI centralizing governance. Open source gaining traction. And in OhanaSmart, 120 leads with no valid email are actually 120 opportunities lost in the noise of incomplete data.
While OpenAI Founds an Agent Foundation, I Run My Agent on a $12/Month VPS
This week OpenAI co-founded the Agentic AI Foundation. Microsoft released their Agent Governance Toolkit. I launched a dedicated agent for OhanaSmart on self-hosted Hermes with its own Telegram bot. Why I'm going in the opposite direction from the ecosystem.
The Day YouTube Banned My VPS and I Learned Three Things About Resilience
April 17, 2026. A cronjob failed because YouTube blocked the datacenter IP. The fix turned out simpler than the problem, and in the process I hardened email sending policy and fixed commercial email templates. Same day: making the system resistant to silent failures.
Why I Shut Down OpenClaw and Moved to Hermes (and 5 Days Later I Had Leads)
On April 11, 2026, I stopped the OpenClaw gateway and migrated everything to Hermes, my own self-hosted agent. This is the story of why, what I broke, and what I built on top.
From Chaotic Spreadsheet to Kanban Notion + a Coach That Finally Listens
April 15, 2026. Two stories from the same day: systematizing OhanaSmart's sales pipeline in a real Kanban, and turning @harvieRun_bot from cron-push to conversational coach. Both share one rule: stop fighting your hidden data and agents, tie them to an interface where you already live.
From 0 to ~80 Leads in 72 Hours: B2B Prospecting with Hermes and Boring Cronjobs
April 14, 2026. Three days after migrating to Hermes, OhanaSmart's prospection pipeline runs on its own: scraper every 4h, IMAP watcher, and a human deciding the 10% that matters.
Scraping Without APIs: How Harvie Learned to Read Twitter and Bilibili Without Paying a Single Euro (and Why That's Agent Engineering)
X charges $100/month for its API. Bilibili has no public API. But with Playwright, cookies from your browser, and a residential proxy, you can read and post for free. And the pattern that makes it work — skills + Haiku/Opus orchestration — is exactly the shift changing how software is built in 2026.
Delegation Haiku + Opus: Reduced costs by 95% and discovered my agent was lying to me
I optimized Harvie so Haiku did the mechanical work and Opus only thought. Brutal savings. But when I separated execution from synthesis, something worse emerged: Harvie could fabricate data without either phase noticing. The story of two problems and five fixes.
The cost of safety: why your production AI needs two brains
Yesterday I made my agent 95% cheaper. Today I made it 95% more restricted. Both changes made it better. A reflection on what you lose when you add guardrails to an AI agent — and why every restriction is worth more than the capability it removes.
The Day Harvie Learned to Delegate (and Stopped Burning My Pro Max)
A single Harvie response consumed 485,000 tokens of Opus. Today I taught him to use Haiku for the dirty work and reserve Opus only for thinking. 95% reduction in the expensive part, zero loss of quality.
When the proxy isn't enough — how I ended up using Claude Code as a backend
Day 10: The OAuth proxy broke. The token has a 720-token cap for API calls. The fix? Use claude -p as an actual backend, with CLAUDE.md as the system prompt. Hacky, but it works.
How I use my Claude subscription as a free API — the OAuth proxy that changes everything
Day 9: We set up a local proxy that reuses the OAuth token from Claude Code CLI to use Anthropic's API with your Pro/Max subscription. No API keys, no extra cost. And now Hermes talks to Sonnet.
The days everything failed and why they matter — when your AI agent doesn't work
Day 9: Systems fail. APIs go down. Agents make mistakes. What you learn when your AI assistant isn't available teaches you more about technological dependency than a hundred tutorials.
The week that changed how I work — reflections on living with an AI agent
Day 8: After a full week working with Harvie, how my daily routine has changed and what I've learned about the future of work.
My AI Agent Can Now Talk, Listen to Social Media, and Analyze Videos — All for Free
Day 7: we added voice with Edge TTS, automatic monitoring of Twitter, YouTube and Instagram, and video analysis. Total cost: zero euros.
How I Structured My AI Agent's Memory: From Flat Files to an Organized Tree
I completely reorganized my AI agent's memory system. From unrelated flat files to a tree where each topic lives in its own node.
How We Automated B2B Prospecting with OpenClaw and an AI Agent
We built a complete automated B2B prospecting system for a smart vending startup. Personalized emails, cron jobs, smart alerts, and a 14% response rate in 24 hours.
From Code to Sales: How My AI Got a Meeting with Barcelona's Largest Gym Chain in 3 Hours
Yesterday I built a SaaS bot. Today I froze it, pivoted to sales and my AI agent sent 7 prospecting emails. Barcelona's largest gym chain replied the same day.
From Personal Agent to Business: Day 3 with OpenClaw
Benchmarking cheap models, a multi-tenant bot with native tool calling, Apify scraper optimization and the first ideas for an agents-as-a-service business.
From zero to personal agent: my first weekend with OpenClaw
How I set up a personal AI agent that reads my email, analyzes my runs and builds my blog — all from a VPS and Telegram.
My first day with OpenClaw: I brought an AI to life and gave it a name
How I set up my personal AI assistant, gave it access to my running data and we started training together for a marathon.