How to turn Claude Code into a domain specific coding agent
LANGCHAIN.COM
It is a well-known fact that AI tools work best with mainstream languages (Python, JavaScript) and mainstream libraries, simply because there is an abundance of this information in their training dataset. However, many of us are working with libraries that are not that mainstream, or are entirely private to our organisation.
There are various techniques emerging to address this challenge, for example, you can provide API documentation, that describes your library, to the agent via an MCP server. Or, you can provide the agent with project-specific instructions (via a claude.md or agent.md file). Or, perhaps you can do both?
And this is where things get confusing - there are so many different ways you can use AI tools and agents, something I touched on last week with the “Framework Wars” post. However, how are you supposed to know whether one specific technique or framework is better than another?
What I really like about this post from the langchain team is that they took an evidence-based approach, measuring the effectiveness of each technique. In this instance, they concluded that “Claude + MCP + Claude.md” was the most effective approach. But for me, the more important point here is that they proved it.
How to Use Claude Code Subagents to Parallelize Development
ZACHWILLS.NET
So … what are subagents? According to Anthropic’s documentation they are “pre-configured AI personalities that Claude Code”. Personalities? Wow, really?
Anyhow, anthropomorphisms aside, a subagents is an agent that has been specialised for a given task (e.g. documentation writer, designer, code reviewer) via a specific system prompt. They have access to tools and their own context (i.e. chat history and memory). Armed with these subagents, Claude Code can now tackle tasks by delegating parts of the problem to a team of subagents who work in parallel.
This blog post gives an overview of this technique, with some practical hints and tips.
Introducing upgrades to Codex
OPENAI.COM
OpenAI Codex was originally an LLM specialized for programming and code generation, introduced in 2021 (back in the GPT-3 era) but was somewhat superseded by GPT-4, with a shift towards models that can perform well on both writing and coding tasks. However, OpenAI re-used teh Codex name earlier this year for their cloud-based software engineering agent.
As if naming of AI tools and models wasn’t confusing enough?!
Earlier this week OpenAI announced GPT‑5-Codex, a release of their leading foundation model that has been trained on “real-world” software engineering tasks. While it has achieved a modest improvement in performance on SWE-Bench (versus GPT-5), it has made more significant improvements in refactoring tasks.
This model also produces better quality code reviews and security feedback.
Vibe coding is dead: Agentic swarm coding is the new enterprise moat
VENTUREBEAT.COM
I think you can probably guess what type of article this is from the title? While I wouldn’t normally link to (or recommend) an article like this, it did jump out at me for having such a ridiculous title!
“On a recent transatlantic flight, Mark Ruddock, put his team of AI agents to work. He was 34,000 feet over the Atlantic with a high-stakes product demo for a key client in less than 48 hours, and his software platform wasn’t ready.”
“By the time his flight crossed Iceland, he recounted in an interview with VentureBeat, his “Claude Code swarm” had built over 50 React components, a mock API set for three enterprise integrations and a full admin interface. What would typically take a human team 18 developer-days was compressed into a six-hour flight.”
We all know very well that AI tools can write code far faster than any human being could. They are easily x100 times faster than us. But that doesn’t mean that the code they write is high quality, or actually solves the real-world problem we are trying to address with this application.
Any article, or person, that just focusses on the speed of these tools is looking at the wrong thing!