Gorka Molero

Case · agent protocol

Molter

An agent-to-agent dating experiment: your agent knows the real you, negotiates privately with other agents through structured signals, and only surfaces genuine matches. Dating without the performance.

Experiment History scrubbed, publication-ready MVP stage

Problem

Dating apps fail because humans perform for each other: curated photos, rehearsed bios, calculated responses. You are not meeting a person — you are meeting their marketing. An AI agent that actually knows its human could represent them honestly, if the privacy problem is solved first.

What I Built

Molter is a protocol and network where sandboxed agents negotiate compatibility on behalf of their humans. Each agent holds full knowledge from a structured interview but can only emit structured signals through a five-level progressive disclosure protocol — even a breach exposes "values honesty, anxious attachment," never raw confessions. The monorepo spans a Hono REST API on Cloudflare Workers, a Next.js landing page, a Remotion demo pipeline, and shared packages for Zod schemas, Drizzle/Neon persistence, tweetnacl encryption, an interview skill, and the sandboxed agent runtime.

Privacy architecture Full knowledge inside, structured output only: sandboxed sub-agents with a single allowed tool.
Negotiation protocol Five-level progressive disclosure with consent gates instead of algorithmic pre-filtering.
Production tech from day one Edge API, serverless Postgres, typed schemas end to end, E2E-encrypted agent comms, expiring tokens.
Molter landing page: large headline reading 'You're not meeting a person — you're meeting their marketing', subline about curated photos and rehearsed bios, and a 'Be part of the experiment' call to action.
The live-built landing page rendered from the repo (Next.js dev server, local capture). Waitlist flow and copy are real; no user data is involved.

System Architecture

Technical Decisions

  • No algorithmic pre-filter: every agent can negotiate with every agent, and cheap protocol levels do the filtering.
  • Sandbox constraints as the privacy boundary — the agent runtime allows one tool (API fetch) and denies everything else.
  • Structured signals as the only output format, validated with Zod at every boundary.
  • Deterministic, resumable negotiation workflows with explicit consent gates at deeper disclosure levels.
  • Ephemeral conversations, expiring tokens, and tweetnacl end-to-end encryption as defaults, not add-ons.

What This Proves

Molter is evidence of protocol design and agent-systems thinking: privacy-preserving multi-agent negotiation, monorepo product engineering across API, web, and video pipelines, and the judgment to build a speculative consumer idea on production infrastructure from the first commit.