Every Vibe Code Camp
Talk 04

Infinite loop agent

Ryan Carson
ralph-loop

Summary

Overview

Ryan Carson, founder of Untangle and former builder-in-residence at AMP, presented his open-source "Compound Product" system - an autonomous loop that analyzes business data, identifies problems, creates PRDs, and implements fixes without human intervention. The session demonstrated how to evolve from manual product iteration to fully autonomous product improvement cycles running 24/7.

Main Discussion

The Problem: Human Bottleneck

Ryan identified the core inefficiency in current product development: founders and builders are the bottleneck. They must:

  • Analyze data manually
  • Decide what problems to fix
  • Write requirements
  • Kick off implementation

    • This means your product only improves when you're actively working on it.

    The Solution: Compound Product Loop

    Building on Kieran Klassen's "Compound Engineering" concept and Jeff Hunt's "Ralph Loop," Ryan created an automated system that runs as a cron job:

    Phase 0 - Daily Intelligence Report
    • Cron job pulls data from database (Postgres)
    • Sends to Opus 4.5 with prompt: "What's the one thing we should fix?"
    • Generates markdown report with actionable insights
    • Delivered via email every night
    Phase 1 - Analysis
    • Bash script reads the daily report
    • Creates analysis.json with prioritized suggestions
    • Identifies specific problems (e.g., "free trial signup conversion is broken")
    Phase 2 - Planning
    • Creates feature branch
    • Loads PRD skill to write detailed requirements
    • Creates atomic user stories with clear acceptance criteria
    • Uses agent-browser for browser testing in acceptance criteria
    • Outputs prd.json with tasks marked complete: true/false
    Phase 3 - Implementation Loop
    • Picks task from JSON
    • Implements the task
    • Runs quality checks
    • If fails: fixes; if passes: commits
    • Updates progress.txt (short-term memory)
    • Updates AGENTS.md if learning should persist (medium-term memory)
    • Loops until all tasks complete
    Phase 4 - Ship
    • Pushes branch
    • Creates PR
    • Founder wakes up to implemented feature

    The Power of Bash Scripts

    Ryan emphasized why Bash is central to this approach:

    • Can invoke agents (AMP, Claude Code) non-interactively from command line
    • No context limits when spawning new instances
    • Can run thousands of parallel instances
    • Simple orchestration without complex frameworks

    Memory Architecture

    The session revealed a three-tier memory system:

  • Long-term memory: Trained into the model
  • Medium-term memory: AGENTS.md file (persistent learnings)
  • Short-term memory: progress.txt file (session-specific gotchas)

    • Key Takeaways

    • You are the bottleneck - If your product only improves when you're working, you're doing it wrong
    • Daily reports are table stakes - At minimum, have a cron job generating AI-analyzed business insights every night
    • For 15 cents you can have a VP - Opus 4.5 can act as VP of Product, Engineering, or Marketing analyzing your data daily
    • Opus 4.5 is the enabler - The model is finally good enough to complete any well-defined, atomic user story with clear acceptance criteria
    • Think in loops - Humans are loops (wake, read, decide, implement, sleep). Agents should be too.
    • Bash is back - Simple Bash scripts avoid context limits and enable massive parallelism
    • Acceptance criteria must be verifiable - Use tools like agent-browser so agents can actually check their own work
    • Three tiers of memory - Long-term (model), medium-term (AGENTS.md), short-term (progress.txt)
    • Compound Engineering + Compound Product - Layer product loops on top of engineering loops for autonomous improvement
    • FOMO is real - If you're not running multiple agent instances 24/7, you're getting out-iterated
    • Start simple - Pull data from Postgres, throw it in a prompt. Don't overcomplicate.
    • Features should solve business problems - Not just technical tasks, but things identified from actual data
    • Agent agnostic - These patterns work with AMP, Claude Code, Cursor, Codex
    • Record your brain state - Journal your thoughts so agents have founder context (Kieran's Mac Mini approach)
    • Open source and iterate - Submit PRs, make it better together

    Memorable Moments

    The Realization Moment

    "I was reading this report the other day and I'm like, why, why don't I just automatically do whatever it's suggesting? This is kind of dumb that like it's waiting for me to say, yeah, go do it."

    The FOMO Tweet

    "I have massive FOMO all the time now because I don't have three instances of AMP running 24/7. Like, and it's real. The speed that we're all going to get out-iterated here is unprecedented."

    The AGI Paradox

    "I thought AGI, which it feels like we basically are sort of here, I thought it would make my life calm but it's made it very stressful."

    The Kieran Roast

    When Kieran asked why he doesn't feel the need for Ralph loops:
    "I think you're not shipping big enough features, Kieran. Sorry."

    Humans Are Loops

    "All we as humans are, are loops. Like you wake up, you read your email, you look at your data, you decide what to do, you implement it and you go to sleep, right? This is what our armies of agents should be doing."

    The VP for 15 Cents

    "This is why you hire people. We used to hire VP of marketing because they would look at data every day and they would give you actionable insights. Well, it turns out for 15 cents, you can get that every day."

    Key Concepts

    Compound Product Loop

    An autonomous system that continuously improves your product by analyzing data, identifying problems, planning solutions, and implementing fixes without human intervention. It combines daily intelligence reports with automated PRD generation and task execution loops.

    Ralph Loop (Ralph Wiggum Loop)

    A Bash-based execution pattern where an agent loops through a JSON list of tasks, implementing each one, running quality checks, fixing failures, committing successes, and updating task status until all are complete.

    Compound Engineering

    A practice of having AI agents learn from their work and persist valuable learnings to the codebase (AGENTS.md), creating compounding knowledge over time. Attributed to Kieran Klassen.

    Three-Tier Memory Architecture

    A memory system for agents consisting of: - Long-term memory: What's trained into the model - Medium-term memory: The AGENTS.md file with project-specific learnings - Short-term memory: progress.txt file with session-specific gotchas

    Daily Intelligence Report

    An automated cron job that pulls business data from your database, sends it to a frontier model (Opus 4.5), and generates a markdown report identifying the single most important thing to fix.

    Notable Quotes

    "The problem with that is that you are now the bottleneck. So you need to be thinking of the ideas to improve your app and then proactively going and building those things. - Ryan Carson (01:34:49)"
    "I was reading this report the other day and I'm like, why, why don't I just automatically do whatever it's suggesting? This is kind of dumb that like it's waiting for me to say, yeah, go do it. - Ryan Carson (01:36:40)"
    "I shouldn't have to wake up and decide what got built overnight. Like I should wake up to a PR, right? - Ryan Carson (01:46:36)"
    "What Compound Product does is basically does this in a loop automatically every day. - Ryan Carson (01:35:02)"

    Tools Mentioned

    AMPClaude CodeCursorCodexFactory / DroidOpus 4.5 (Claude)GPT 5.2agent-browserBashPostgresCompound ProductCompound EngineeringRalph Loop / Ralph Wiggum LoopCora (Every's)Monologue

    Transcript

    04 of 18
    ← Previous Personal AI suite Next → AI consulting workflows