Don't Waste Your Back Pressure

Summary

Moss Ebeling argues that successful AI agent applications rely on back pressure—automated feedback mechanisms that help agents identify and correct mistakes during task execution. Engineers gain leverage by building these feedback systems rather than manually reviewing each output.

The Core Problem

Without automated feedback, agents depend on human review: "Without a way to interact with a build system the model relies on you for feedback about whether or not the change it made is sensible." Manual correction doesn't scale.

Categories of Back Pressure

Build Systems & Bash Tools

Agents running their own builds and reading error output eliminates trivial feedback tasks. The build system becomes an automated reviewer.

Type Systems

Languages with expressive type systems—Rust, Elm—prevent invalid states and generate helpful error messages. These messages guide LLMs toward correct solutions without human intervention.

Visual Rendering Tools

MCP servers for Playwright and Chrome DevTools let agents compare UI expectations with actual rendered results. Visual verification without manual inspection.

Formal Verification

Proof assistants like Lean combined with AI enable trusted results through automated validation. The tightest possible feedback loop.

Testing & Fuzzing

Randomized testing for CUDA kernels and automated documentation generation provide immediate correctness feedback.

Diagram

graph TD
    Agent[AI Agent] --> Output[Code/Output]
    Output --> BP{Back Pressure Systems}
    BP --> Build[Build System]
    BP --> Types[Type Checker]
    BP --> Visual[Visual Renderer]
    BP --> Verify[Formal Verification]
    BP --> Test[Tests & Fuzzing]
    Build & Types & Visual & Verify & Test -->|Errors| Feedback[Error Signal]
    Build & Types & Visual & Verify & Test -->|Pass| Next[Next Task]
    Feedback --> Agent

Key Insight

The central question: "Are you wasting your back pressure?" Rather than spending time correcting agent mistakes, invest in automated quality checks. Shift focus from trivial corrections to higher-level problem-solving.

Connections

• dont-waste-your-back-pressure - Geoffrey Huntley explores the same concept with a focus on calibration—balancing too much vs. too little backpressure
• essential-ai-coding-feedback-loops-for-typescript-projects - Matt Pocock's practical implementation of feedback loops using TypeScript, Vitest, and Husky
• context-efficient-backpressure - Complementary focus on making backpressure context-efficient by suppressing verbose output