Multi-Agent Strategies

Use multiple AI models for different aspects of spec creation, review, and building. Supaspec tracks every change with full agent attribution, so you always know who wrote what.

Why Multiple Agents?

Different models excel at different tasks. Using them together produces better specs and builds:

Deep Architecture (e.g., Claude Opus)

Best for initial spec generation, complex data modeling, and system design decisions that require deep reasoning.

Quick Review (e.g., Claude Sonnet)

Fast review passes to catch inconsistencies, suggest improvements, and validate completeness. Great for proposals.

Specialized Agents

Security audit LLMs, market research agents, performance specialists — each contributing to relevant sections.

Proposals for Agent Review

When a second agent reviews work by a first agent, it should use propose mode instead of direct commits. This creates a proposal (like a pull request) with line-by-line diffs that you can accept or reject.

This prevents agents from overwriting each other's work and gives you full control over which changes land. See Proposals for details.

Agent Attribution

Every commit in Supaspec records which agent (or human) made the change. The version history shows:

  • Agent name and color-coded badge
  • Commit message and description
  • Timestamp and version number
  • The prompt that triggered the change (if provided)

This makes it easy to compare outputs from different agents and decide which approach is best. See Agents for how agent identity works.

Recommended Workflows

Generate → Review → Build

  1. Use a powerful model (Opus) to generate the initial spec
  2. Use a fast model (Sonnet) to review via proposals
  3. Accept/reject proposal changes, then approve sections
  4. Use a coding agent (Claude Code, Codex) to build from the approved spec

Parallel Specialized Review

  1. Generate the initial spec with any model
  2. Connect a security-focused agent to review auth and data sections
  3. Connect a UX-focused agent to review the UI layout section
  4. Connect a performance agent to review the data model and API spec
  5. Merge the best suggestions from each review
← Assets & SecretsMCP Tools Reference →