Backpressure for Coding Agents: Let Machines, Not Humans, Catch the Slop

Lucas Costa argues that the two default modes of using coding agents are both broken: turning them loose produces unreviewable PR floods, while babysitting every keystroke negates the point of delegation. The missing third option is borrowed from systems engineering — backpressure. Just as a slow consumer forces a producer to slow down or buffer, automated checks should refuse an LLM’s work until it meets the bar, instead of humans serving as the soft, slow validator at the end of the pipeline.

The historical analogy is TypeScript, linters, and CI: each one shifted correctness checks from reviewer eyeballs to machines that reject bad work at the boundary. With LLMs, that shift has stalled. Developers now act as expensive clipboards, ferrying review-bot output back into the coding agent, because the guardrails haven’t been wired into the agent’s loop. Costa’s fix is to stack mechanisms — lint, tests, verification scripts, manual cURL and browser checks, benchmarks, specialized review agents for functionality/tests/types/brevity, plan reviews, visual design reviews, and PR monitoring — so the model has to satisfy them before declaring a goal done.

He packages this as a Claude skill (npx @lucasfcosta/backpressured, invoked via /backpressured) that iterates against a goal while running the checks, customizable via a BACKPRESSURE.md file. The broader point: longer unattended agent sessions only become safe and useful when the machinery around the agent — not the human reviewer — is doing the low-level quality enforcement, freeing humans to weigh in on design and complexity instead.