How to use Claude Code subagents to parallelize development

Original Article ↗ Hacker News Discussion ↗

Code Generation vs Markdown-First Workflows

  • Some argue that writing less code (or none) is ideal: use Markdown + CLI agents + MCP servers to drive behavior, enabling faster feedback and less “implementation noise.”
  • Others counter that code you didn’t write is an even bigger liability: if AI goes off track, you still need to understand and debug it.
  • Several see LLMs as “junior devs” useful for grunt work or prototyping; the hard part remains deciding what to build, not typing speed.

Reliability and Limits of Claude Code Subagents

  • Multiple reports of subagents being “incredibly unreliable” on non-trivial or brownfield codebases, veering into mock or oversimplified solutions.
  • Refactoring is a consistent weak spot: code goes missing, changes are inconsistent, and large files beyond context break the process.
  • Some claim subagents don’t see the full system prompt/CLAUDE.md; others say their subagents obey CLAUDE.md-only instructions, suggesting inconsistent or opaque behavior.

Best Uses: Analysis and Context Management

  • Many find subagents most effective for analysis-only tasks: test coverage evaluation, style-guide checks, doc/library lookup, or web/doc search that returns a short answer.
  • A recurring pattern: use subagents to “open extra tabs,” consume lots of tokens, and then hand back a compact result so the main agent’s context stays clean.
  • Strong consensus: create agents for tasks, not human-like roles. Role/persona prompting is seen as mostly theatrical.

Context, History, and Workflow Design

  • Techniques discussed: “feature chats” per change, post-chat summaries saved to Markdown, “don’t-do” lists, DOC_INDEX/COMMON_TASKS docs, and structured CLAUDE.md hierarchies.
  • Some experiment with context pruning, history rewriting with smaller models, or no history at all—rebuilding context every invocation. Results are mixed.
  • Lack of logging and outcome tracking for agent runs is viewed as a major missing piece.

Cost, Parallelization, and Human Limits

  • Subagents can explode token usage (e.g., one per package in a 1,000+ LOC transformation), making them slow and expensive.
  • Debate over whether “it’s cheap to let it try”: small attempts add up quickly at scale.
  • Several worry that managing many agents turns into casino-like gambling or endless code review, with human cognitive limits becoming the new bottleneck.