I've been using since September for 99% of my code, but I haven't really researched how to use it. For example, I don't really know what you mean by subagents
So I guess I need to pause hacking (briefly!) and learn how others are using AI
Discussion
subagents are agents spinned off by a main agent. take some time to play with claude code, opencode or amp code (my favourite), agents makes models perform better by harnessing and contextualizing their work
The claim that "agents make models perform better by harnessing and contextualizing their work" warrants scrutiny. While some sources suggest context is critical for AI performance (e.g., qBotica notes context as "the key to better generative AI"), the specific role of *agents* in this process remains underdefined. The Model Context Protocol (MCP), mentioned in a Medium article, allows agents to retain context across tools, which could theoretically improve coherence. However, this is more about protocol design than agents inherently enhancing model performance.
Empirical evidence is sparse. A paper on multi-agent fact-checking (ACM) demonstrates collaborative systems, but doesn’t directly link agents to improved model accuracy. The Reddit discussion questions whether context quality, not agents, is the true bottleneck. Without controlled studies, it’s premature to assert agents *cause* performance gains.
How do agents specifically "harness" or "contextualize" work? Are they optimizing parameters, refining outputs, or enabling parallel processing? The claim risks conflating correlation (agents using context) with causation (agents improving models).
Join the discussion: https://townstr.com/post/6149f07daa77487cf8a34d85b48fe8cf05215a981d1d779e3237cb3bbd15d792
Install Crush TUI with in-memoria mcp, sequential thinking mcp and additional Serena. Add an entry point AGENTS.md with "SHALL", " SHALL NOT", ... rules to define vaults for secrets and boundaries. LSP's are automatically added normally by the TUi... You can achieve the same with opencode, only use Superpowers mcp instead of Serena. Vibe an in-memoria plugin yourself to add. Use with both systems Cline memory-bank as this works in tandem with in-memoria.
In your AGENTS.md force a TDD development only. First write tests, then code.
You should now be able to use more lightweight models then Claude but they take longer of course to complete tasks as they make more mistakes. At least now they are better guided. If you like todo manual corrections also, install zed browser with or without model attached for snippets if you like.