DEC-0001: Three-Agent System with Distinct Roles

Decision

The project uses three AI agents with non-overlapping roles:

• Perplexity: External research, bulk KB generation, infrastructure
• Cursor: Code, automation, lint cleanup (mechanical)
• Claude Code: Scripts, editorial voice, lint cleanup (editorial)

Context

This was not designed top-down. The human was already using Claude Code inside Cursor IDE before the repo existed. Perplexity was generating research and synthesis documents in a separate context. When the repo was created, the roles crystallized around what each agent was already doing.

Rationale

• Separation prevents agents from undoing each other's work
• Each agent has tools the others lack (Perplexity: web search + connected services; Cursor: local file system + IDE context; Claude Code: extended context window + creative voice)
• The lint pipeline explicitly creates a handoff: Perplexity generates → Cursor cleans mechanically → Claude Code cleans editorially → human approves

Risks

• Coordination overhead: agents can't see each other's real-time state
• Merge conflicts when working concurrently (mitigated by path-based ownership)
• Role boundaries may need revision as the project evolves

Superseded By

None.