I break down how to choose between GPT-5.6 models (Luna, Terra, Sol) and reasoning levels through a simple analogy with human attention: some tasks need more "memory for rules," others need more "time to think."
Explaining Kent Workflows - a new paradigm for working with agents that lets you turn manual agent management into a fully automated process where agents work in loops.
Released Kent 2.0 - formerly Builder, now Kent with a new site at kent.sh. This release includes: system prompt customization, new subagents with config examples, performance and terminal fixes, system sleep prevention, and more.
Firefox CLI is a CLI interface for controlling a real Firefox session with AI agents - a full equivalent of Agent Browser for Chrome, but better. Install once via an extension; supports Claude Code and Codex.
MCP has fundamental architectural problems - it pollutes context, kills flexibility, and creates security risks. For agents with terminal access, CLI is a far better alternative, and it's time to stop using MCP.
Complex systems of agents, slash commands, and orchestrators aren't progress - they're micromanagement that gets in the way of frontier models doing their best work. Here's why the simple approach works better, and how I actually work.
Explanation of Kent's advanced compaction algorithm and how it lets the agent run for dozens of sessions in a row without losing context or making unnecessary stops.
Introducing Kent - a free, open-source coding agent built for professional engineers. Outperforms Codex and is already available on GitHub.
Google released Gemini 3.1 Pro with fixed tool calling, Anthropic and Google started aggressively banning subscription use outside their own tools, while OpenAI keeps actively supporting OpenClaw and is gearing up for GPT-5.3.
Spent 5 days and $150+ testing OpenClaw with different AI models. GPT is a robot without personality, Gemini has great character but can't work properly, Opus is perfect but costs $100/day. OpenClaw is still full of bugs (4000 issues, 3000 PRs).
OpenClaw (formerly ClawdBot) by Peter Steinberger is a personal AI assistant with file-system access, background agents, and remote control. This post covers its origins, hype, and my hands-on take.
Agent skills are a way to progressively disclose context to AI through markdown files, which saves tokens and boosts performance. I explain why you need them, when to use them, how to create them, and debunk the myths.
In December, AI models crossed a threshold after which I became useless as a programmer. Now I push code to prod that Ive never read.
Introducing ksrc - a CLI utility that simplifies searching and viewing Kotlin library source code from 15+ commands down to two for your AI agents.
Detailed comparison of Codex and Claude Code after three months of using both tools. Breaking down the pros and cons of each, model features, harnesses, pricing, and revealing who I paid $200/month to
Anthropic published research showing that large language models are capable of detecting manipulations of their thinking process and recognizing their own intentions, which may indicate elementary forms of self-awareness.
I built a complete blogging platform with automated publishing across multiple languages and channels in 8 hours, using Claude without writing a single line of code myself.
Claude 4.5 Sonnet is a breakthrough update that completely replaces Opus for most tasks thanks to better instruction following, parallel tool calls, and impressive speed. You can now work with lower limits and greater productivity.