Long-running agents fail less from prompts and more from harnesses, says a Reddit engineer

Confident_Analysis89 · reddit · 2026-07-24

The author argues that once an agent runs for hours, most failures stop looking like prompt problems and instead come from the surrounding harness: stale docs, copied bad patterns, context bloat, weak validation, and poor stopping conditions.
They highlight “harness engineering” as the real lever: keep AGENTS.md small and use it as a map, move critical rules into linters and runtime checks, and rely on hard evidence instead of hoping the prompt will be remembered.
The evaluator-agent idea is still unresolved. In one experiment, splitting planner, generator, and evaluator produced a launchable app, but the core behavior was broken and the setup cost about 20× more.
Their current takeaway is to start with hard checks and real runtime evidence, then add separate evaluators only for cases normal tests cannot judge.

More from coding & agent