Notes from my AI practice: agentic coding, multi-agent architecture, mid-market use cases.
An email assistant for macOS Mail.app that runs on-device: PII redaction before every LLM call, and a silent macOS 26.5 bug that quietly killed the fast SQLite query.
Three agents running around the clock: Sven scouts AI ideas, BuilderBob runs agenticbuilders.at with 19 cron jobs, session-orchestrator drives the coding loop. Architecture, hard escalation gates, and what agentic engineering actually looks like in practice.
An open-source multi-agent tool matures in 9 days across 18 sessions: from the /plan to /evolve lifecycle across four runtimes. What daily self-use made of it.
Qwen3.6-35B on an M4 Pro 24 GB: technically loaded, 100% GPU – and yet 0.1 tokens per second. An honest diagnosis, five lessons, and hardware recommendations for others.
After two months and 1,052 commits, I swapped the AI layer in my autonomous infrastructure system from OpenClaw to Hermes Agent by Nous Research. Why the learning loop made the difference.
Two weeks after my first hackathon: 1st place at BitGN PAC 2026 at AI Factory Austria in Vienna. Autonomous AI agent, 104 tasks, 5 security layers, blind scoring.
After 3,000+ sessions with Claude Code: What systems prevent AI agents from drifting in long-running projects. CLAUDE.md, Rules, Session Handovers and Quality Gates.
An honest week as a solo founder: GitLab migration, n8n automations, a live content pipeline on Instagram, and the question of how far AI agents can really go today.