Developer advocates automated 'backpressure' to improve AI coding agent reliability
A new technical analysis and accompanying npm tool, @lucasfcosta/backpressured, integrate strict type checking, continuous integration tests, and automated review agents to validate code before human intervention.
Software engineer Lucas Costa has published a technical analysis titled "Backpressure is all you need," arguing that human review acts as an inefficient bottleneck for AI coding agents. Costa identifies two flawed approaches in current workflows: letting large language models run unattended, which risks bugs and review floods, or treating them as restricted autocomplete, which is slow and defeats the purpose of delegation. To resolve this trade-off, he proposes implementing automated 'backpressure' mechanisms to validate code before human intervention.
In systems engineering, backpressure forces a producer to slow down or shed load when a downstream component cannot accept more work. Costa applies this concept to software development by building mechanisms for the agent to validate its own work. The goal is to make longer unattended sessions safe and useful while reducing the number of low-quality pull requests that teammates must review. This approach aims to move the "no" off humans, allowing them to focus on higher-level feedback rather than low-level correctness.
The proposed solution involves a suite of automated checks integrated into the development loop. These include automated tests and linting to catch correctness issues early, as well as benchmarking suites for performance-sensitive applications to detect regressions. Costa also implemented manual testing skills taught to the model via the obra/superpowers builder, enabling the agent to verify local application behaviour. Additionally, review agents were introduced to check for subjective quality issues such as readability, complexity, and loose types.
Further layers of validation include planning phase reviews, where a subagent checks the fundamental approach before implementation begins, and pull request monitoring to detect conflicts or continuous integration failures after the pull request is opened. Costa also explored visual design reviews using Playwright MCP to compare screenshots against design files, though he notes uncertainty regarding the efficacy of this specific mechanism.
To operationalise these concepts, Costa has released a tool, @lucasfcosta/backpressured, available via npm. Developers can invoke the tool in Claude using the /backpressured command or by creating a custom BACKPRESSURE.md file. The tool integrates these checks into the development loop, allowing agents to work unattended more safely. Costa remains open to refining the approach, including breaking down review agents into multiple specialised agents to improve reliability.
