Practical insights on AI agents, Claude Code, and automation. Written by Pawel Jozefiak from 1000+ real sessions.
Full articles on Digital Thoughts. Summaries here are optimized for quick reference.
A practical beginner guide to using Git and GitHub as a checkpoint system when AI is editing your project. The useful habit is simple: commit before the agent changes anything serious, keep secrets ig...
A reminder that building your own tools is not only about saving money or replacing SaaS. It teaches you the shape of the problem, gives you judgment, and makes future AI-assisted work better because ...
The question is not how capable the agent is. The question is what capacity you give it. Runtime, cost, file scope, authority, and escalation rules matter more than vague autonomy. Useful agents are b...
A local model experiment can turn into a resource problem fast. The lesson for agent builders is to treat CPU, memory, swap, temperature, and model size as operating limits, especially for unattended ...
No single model or harness wins every task. Opus 4.7 changed the cost/performance tradeoff enough that Codex became useful again as part of a split workflow. The practical lesson is to route work by t...
When a model gets more expensive for the same workflow, hidden waste becomes visible. File re-reads, failed web fetch loops, retry spirals, and context thrash are not harmless annoyances. They are bud...
A practical comparison of AI coding harnesses in real autonomous workflows. The important finding is not that one tool wins. It is that each harness has a job: planning, repo edits, reviews, interacti...
A beginner-friendly agent architecture guide: start with a folder, an instruction file, a memory layer, and progressive permissions. The common mistake is trying to make the first agent too autonomous...
Claude Code version 2.1.88 accidentally shipped a 59.8MB source map file inside its npm package. A missing line in .npmignore exposed 1,900 TypeScript files and 512,000 lines of Anthropic's internal o...
AI agents make you dramatically more productive, but the productivity gains do not translate into more free time. In two months, I shipped 16 digital products, maintained a daily blog, and ran automat...
Claude Cowork (formerly Claude Computer Use) is Anthropic's attempt at a desktop agent that can see your screen and click things. After extensive testing, the answer to "is it an agent yet?" is: almos...
Letting an AI agent operate with full autonomy reveals failure modes you cannot predict from testing. Real production failures include: the agent sending duplicate messages because it could not detect...
March 2026 is a turning point in AI development tools. Google AI Studio now runs Gemini models directly in the browser with impressive speed. Anthropic released Claude with extended thinking and 1M co...
Giving an AI agent real money and a shopping task reveals the current limits of agentic commerce. With $25 and a virtual debit card, my agent successfully browsed stores, compared products, and got to...
Teaching an AI agent to think means building a self-improvement loop: the agent logs errors, extracts lessons, detects patterns, and adjusts its own behavior. The core architecture has four components...
Migrating an AI agent to dedicated hardware (Mac Mini M4) takes about 72 hours and reveals problems you cannot anticipate on a shared machine. The biggest surprise: headless macOS breaks UI automation...
The difference between an AI tool and an AI agent is identity persistence. My agent has a layered memory system that stores who I am (preferences, patterns, relationships), what I care about (projects...
Digital Thoughts hit 1,000 subscribers in March 2026 after consistent publishing about AI agents and automation. The growth came from three sources: organic search (articles ranking for "local LLM", "...
A local AI agent that runs entirely on your Mac and iPhone is now feasible with Apple Silicon and MLX. The project (Familiar) uses on-device models for routine tasks (message triage, calendar manageme...
The AI adoption gap is real: 88% of companies report using AI in some form, but only 6% see measurable business results. The problem is not the technology. It is implementation. Most companies adopt A...
After testing Google AI Studio, Cursor, Claude Code, and other AI development tools extensively, Cursor remains the best general-purpose AI IDE for most developers. Google AI Studio impressed with its...
Running local LLMs on consumer hardware is now practical for many tasks. Using MLX on a MacBook Pro M3, Qwen 2.5 32B runs at usable speeds for code generation and analysis. On iPhone, smaller models (...
Vibe coding (letting AI build whatever it wants) produces technically functional but creatively dead output. After weeks of daily AI-built experiments, the pattern was clear: AI defaults to safe, gene...
Submitting a hackathon project built entirely by an AI agent overnight proves that autonomous development is production-ready for certain types of projects. The process: I defined the requirements at ...
The best way to structure CLAUDE.md is with a layered system: identity at the top, behavioral rules in the middle, and reference paths at the bottom. After 1000+ Claude Code sessions, I settled on a p...
Running an autonomous AI agent costs approximately $400/month (API calls, hosting, tools). In February 2026, my agent generated $355 in revenue through product sales, content publishing, and automated...
The AI industry shows classic bubble characteristics: inflated valuations, companies pivoting to "AI" without substance, and a gap between hype and real utility. But unlike the dot-com bubble, the und...
Nami is a keyboard-driven window manager for macOS built entirely with AI assistance. It uses a grid-based system inspired by tiling window managers but adapted for macOS conventions. The project went...
OpenClaw is an open-source AI agent framework that does many things right: clean architecture, good defaults, and a sensible skill system. But after evaluating it against my custom-built agent (Wiz), ...
An AI agent night shift is a scheduled automation session that runs while you sleep, executing predefined tasks like content publishing, code reviews, and data processing. The architecture uses macOS ...
After two months of daily use with both Claude Code and OpenAI Codex, the differences are clear. Claude Code excels at autonomous multi-file tasks, understanding project context, and working as an age...
Testing Windsurf, Cursor, and Claude Code side-by-side for real development work reveals that each tool has a distinct sweet spot. Cursor is the best general-purpose AI IDE with the smoothest editing ...
WIZ is a personal AI agent built on Claude Code that runs 24/7 on a dedicated Mac Mini M4. It handles automation, content publishing, code deployment, job searches, morning reports, and overnight task...
Want all 125+ articles? Subscribe to Digital Thoughts.
Subscribe Free