Conversion Factory
Agents that evolve site blocks

May 31 - Jun 5, 2026 · 19 commits · phase 1 complete

spec-first the PRD and the architecture diagram were committed before the agents existed.

An agent runtime for conversion-rate optimization on Edge Delivery Services. A coordinator forms a theory from analytics, four spoke agents produce a variant block on a git worktree, and the change ships as a pull request into a live A/B test. Winners get promoted into the codebase by a separate PR. A retrospective agent then edits the agents' own instructions, so the system learns across cycles.

May 31, 2026 · PRD + architecture diagram, drawn first

Four phases, one diagram

Analyst scores the block. Architect forms the theory. Builder writes the variant. Validator runs Lighthouse, accessibility, and regression.

Conversion Factory architecture: a coordinator hub orchestrates analyst, architect, builder, and validator agents through four phases - generate a PR, human review plus live A/B experiment, promote to permanent baseline, and a retrospective loop that updates agent instructions.

From the repo's ai-docs. Phase 4 is the part I care about: the retrospective agent proposes instruction changes, and humans gate how the agents reason.

id
p-conv
  • 4 spoke agents behind one coordinator
  • A/B baseline vs variant, per visitor
  • 1 false-PASS hole closed: verify wiring, not just pixels

All projects