Martin Thompson

For risky one off refactors search for a codemod or ask the llm to write a codemod with tests first For long running rules ask llm to write lint rules for your language These convert llm output to deterministic functions, guaranteed to not lose learning or context on reruns

3

It's surprising that declarative concurrency isn't more of a thing, given the perils of fiddling about with WHILE loops and low-level synchronisation.

2

2 Comments

Following on from my previous post on the current conflict between Agentic coding suites and GitHub's premium request billing model, I decided to have a go at creating a set of agents that are GitHub premium request friendly. Introducing Squadron: https://lnkd.in/gjjX4K4K It's mainly for my own entertainment, but if it's useful to someone out there then I'm glad it helps. It's based on a spec-driven development concept. You interact with the Refine Requirements agent to define what you want. It asks pertinent questions and creates a backlog of tasks. You interact with the Task Dispatcher agent and tell it to work on the backlog. It then kicks off the suite of agents to analyse, implement, review, test the backlog tasks and ensure the backlog is kept up to date. There's slightly more to it under the hood to keep things interesting 😁 - Multi agent debate approach to code reviews - Context friendly backlog - Review friendly chunks of work

11

For those interested in my AI-built Occam transpiler, this will not mean much to anyone else! I showed the LLM the source for the kroc compiler, summarized the ideas from the previous Claude session (see last week's post) on how to implement support for #INCLUDE of modules, and in less than an hour it came up with this PR: https://lnkd.in/gxxz4FFF One thing I found a bit mind-boggling was that it nonchalantly mentioned in passing that we weren't going to be able to compile all of the "course" module because it used language features that weren't implemented. Interesting because I hadn't asked it to look at the module source, and also interesting because its idea of which language features were implemented came from the source code, not the document that says which features are implemented (it turned out to be incorrect). The kind of thing you'd expect from a very competent human developer: it spots that the general goal it deduced you have, but didn't actually articulate explicitly, can't be achieved for <reasons> that you hadn't asked it to think about. So next we have another iteration to get it to implement those missing language features...

5

2 Comments

Here’s a great simple non-technical explanation of the state of personal AI agents today, such as OpenClaw (Clawbot/Clawdbot/Molt). The evolution from token prediction and tool integration to now providing looping ability to both write and run new tools is both an incredibly powerful accelerant and a bit crazy scary too. As with most things AI-related, the most important thing is to provide it with the most clear and clean detailed context possible to maximize the potential magic of AI while minimizing the negative chaos.

8

Great step-by-step breakdown and an explanation of the latest hype around emerging AI tools. Yes, they are (and can be) incredibly powerful when given proper context, proper instructions and well defined access and blast radius... No, they are not AGI, nor are they sentient, and should not be treated as such.

8

1 Comment

I found this a particularly useful perspective on the evolving discipline growing around AI-assisted software development. Much of the conversation today focuses on the models and tools, but this highlights something I've been seeing more often in practice: success depends just as much or more on the surrounding engineering discipline, context management, evaluation and feedback mechanisms as it does on the AI itself. Well worth a read for anyone trying to separate the hype from the reality of building production systems with AI.

10

1 Comment