What Happens When You Hit Claude’s Limits? | Sam Hesson (Meta AI)

What does it look like when someone truly maxes out agentic coding? Sam Hesson former CTO and co-founder of Nanome, now on Meta’s AI incubation team .He didn’t just push the limits of Claude. He broke them. His $50,000 token spend across two and a half months directly led Anthropic to impose rate limits on their unlimited Claude Max plan.

I sat down with Sam in Seattle to go deep on the custom architecture he built: from talking through feature ideas on a walk with Meta Ray-Bans, to a fully automated CI/CD pipeline that pits parallel agents against each other in a Darwinian battle for the best pull request.

Watch now

In the video:

• The Token Abundance Mindset Why treating tokens as abundant rather than scarce is the foundational shift that unlocks next-level agentic development, and why the math already points toward this becoming the default way of working as token costs continue to fall.
• The Walk Flow: From Ray-Bans to PRD How Sam turns a daily walk in Seattle into a structured product and technical spec, using Meta Ray-Bans with advanced voice mode to brainstorm features, which then feed into custom scripts that auto-generate implementation-aware prompts aligned to his codebase patterns.
• Darwinian CI/CD: Parallel Agents in Competition The core of Sam’s architecture: when a GitHub issue is tagged AI-ready, it automatically spins up three parallel agents on separate DigitalOcean VMs, each tackling the problem independently with its own Dockerized environment and database snapshot, then submits competing PRs.
• LLM Judge and Rubric-Based Self-Healing How a rubric-scored LLM evaluates each competing PR across criteria like documentation, test coverage, and adherence to implementation patterns — with a feedback loop that forces agents to fix deficiencies before the judge selects a winner. Sam only ever reviews the one the judge picks.
• AI Dev Readiness: Preparing Your Codebase Why getting to this level requires upfront investment in proper DB seeding, multiple database snapshots for different testing scenarios, clear auth mocking for different user roles, and a solid Playwright foundation — the groundwork that makes agents reliable rather than chaotic.
• DevPlan: Turnkey Structured Prompting A walkthrough of DevPlan in action, showing how it interrogates your codebase, asks clarifying questions, and generates what Sam calls “juicy coding prompts” — detailed, opinionated specs that dramatically improve agent output quality and adherence to your existing patterns.

💡 Resources

Nanome — Nanome.ai
DevPlan — devplan.com
Bruno API Client — usebruno.com
Sam’s LinkedIn — https://www.linkedin.com/in/samhessenauer/
Sam’s Twitter https://x.com/aquamansam1
One-Shot Podcast — https://aitinkerers.org/podcast
AI Tinkerers — https://aitinkerers.org