Blog

Software agents are not a new concept. In 1966, Shakey the Robot could already reason about its environment and take actions. Modern LLM agents use a very different mechanism, but the high-level autonomous loop of “reason, decide and act” is remarkably similar. The big promise of the current wave was that large language models would create far superior reasoning and tool calling compared to Shakey’s hand-crafted if-then-else trees. If machines could reason using human language, would they be able to reach human levels? ...

Of course, open-source LLMs will win. Anyone who looks at the history of software development can see this trend. ➡️ No vendor lock-in. Qwen 3.6 is the same on every provider, so switching to avoid bad service is trivial. ➡️ Grassroots distribution. Proprietary software needs to convince decision-makers to buy. Open-source targets engineers who convince their boss. ➡️ Data privacy and self-hosting. How long will we keep sending confidential data to companies that benefit from training on it? ...

Imagine a half-marathon race. The runners are pushing themselves, and the crowd is cheering them on. But in that crowd, there are two loud, annoying groups. One of them is belittling the 21K effort. Why run only a handful of kilometres? The human body could run around the globe with the right training and nutrition! Soon, every human will sprint faster than a car! They make up this nonsense in an attempt to sound interesting to their peers. They are not. ...

The all-you-can-eat era of AI is ending. Compute constraints, heavier models, and a fully hooked user base are pushing providers toward pay-as-you-go. That shift will force better choices, smaller models, and fiercer competition between tools.

Running top-of-the-line models ain’t cheap. Burning Opus tokens is expensive. Yesterday, Anthropic sent out an email to all Claude Pro/Max accounts letting them know they are closing the gates for “third-party harnesses”. That’s a polite way of saying OpenClaw users. Anthropic launched Claude Code with a fixed monthly price. They wanted developers to think about what they could build next rather than worrying about what that would cost. But from day one, people have come up with ways to maximise those all-you-can-eat buffets. They wrote infinite loops that kept Claude busy. Anthropic introduced session windows. Users then started setting up cron jobs to maximise their usage, which led to weekly limits. An arms race kicked off. ...

In the last edition of We Should Do Something With AI, I discussed my Daily Doom experiment. How long can a coding agent build software before it all breaks down? I set up a system to autonomously build a Doom clone. Every night, the system would pick up the GitHub issues and add them to the game. If there were no issues, Claude would make something up. New features, new game mechanics. No humans involved. ...

Software experts must balance enthusiasm for generative AI’s genuine capabilities with critical scrutiny to prevent unchecked hype from driving poor technical decisions.

How to manage multiple Claude Code accounts across different machines using Jean-Claude to selectively sync configuration while keeping account-specific data separate.

As software development feedback loops accelerate toward minutes, quality assurance remains the final bottleneck where human judgment is still required.

Generating code is a solved problem. Codex, Claude Code, Mistral’s Vibe and Gemini CLI are all capable of creating good software. But writing code creates a form of entropy. The more we add to a system, the worse it gets. Claude, like a junior developer and their manager, couldn’t care less about refactoring and code quality. All three have no sense of the risk software entropy poses. Seasoned software engineers, on the other hand, are obsessed with refactoring to fight that entropy. ...