Long-horizon coding tasks extend beyond a single code snippet or isolated edit. They can involve planning an architecture, creating several files, running or interpreting tests, generating supporting assets and making follow-up changes without losing earlier requirements.
The main challenge is maintaining coherence over time. An AI system must remember relevant project state, notice when a later change conflicts with an earlier decision, recover from failed attempts and stop when the requested result has actually been reached.
A quick prototype is useful evidence, but reliable evaluation also checks correctness, maintainability, recovery from errors and the quality of later revisions. Long execution alone is not success if the system spends time repeating work or drifting away from the goal.
ELI5
A long-horizon coding task is a software job with many steps that must stay connected. It is more like building a small game than writing one short function because later choices depend on what was created earlier.
For example, an AI might plan the game, create its files, add characters, generate sounds, test the result and then improve one feature. It must keep the original rules in mind throughout the job, and a person still needs to check that the finished project works and is maintainable.
