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

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.