Agents vs. plain code: where do you draw the line??
retained text
I built a due diligence tool for investment deals with Claude Code, and I keep going back and forth on how much should be agents and how much should be regular code. How it works right now (ELI5 version): 1. **I start with a big checklist.** About 1,000 due diligence questions, each tagged with when it applies. 2. **Code narrows it down.** I answer a few yes or no questions about the deal, and a script keeps the roughly 200 questions that fit the deal 3. **An AI agent reads the documents.** It answers each question with the exact quote and page number, or says it is not there, and the question remains open. 4. **Code double-checks the AI.** A script makes sure every quote really appears in the document, then a cheaper AI agent checks that the quote actually answers the question. 5. **Code sorts the answers.** A script marks each answer as proven (signed contract or outside report), claimed (seller’s word or a phone call), or still open, and drafts requests for the important missing answers. 6. **Another AI agent writes the summary.** It drafts the deal overview and memo from the checked answers only, and the documents overall, and a script deletes anything without a source. So code runs the whole process, and the 3 agents only read, check and write. The other approach I have seen uses about 30 agents: a manager agent, phase managers under it, then specialists like a “rent roll analyst” and a “title reviewer”, each with long instructions. Questions: 1. How do you decide what should be an agent and what should be code? 2. For a structured process like this, is “code runs it, a few agents help” the right call, or does a manager-agent setup actually work better? 3. How do you keep agents from getting bloated as things grow? Would love to hear what has worked for you.
document fetched 2026-10-10 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x1nk79/agents_vs_plain_code_where_do_you_draw_the_line