Long-horizon task
A long-horizon task is one in which an agent must maintain progress across an extended sequence of steps. Earlier actions affect later options, and success may be observable only near the end. There is no universal number of steps that separates a long-horizon task from a short one.
How it works
The system may need to remember context, manage intermediate state, choose subgoals, and recover when an action fails. Each step can introduce errors that affect later decisions. Evaluation should inspect the final outcome and relevant side effects, not merely count completed actions.
Computer-use benchmarks and manipulation research illustrate tasks where sequences and compounding errors matter. A short action sequence can still be difficult when observations are ambiguous or mistakes are irreversible.
Why it matters for licensing
Workflow records can be useful when they preserve complete sequences, intermediate states, and outcomes. Isolated screenshots or final records may not explain the dependencies that make the task challenging. Rights and sensitive information must be assessed throughout the sequence.
Example
Fictional example: An agent prepares a draft procurement package by comparing inventory, finding approved suppliers, reconciling quantities, and assembling documents. An early unit-conversion mistake affects later totals, so the evaluation checks the whole result.
Limitations and misconceptions
Longer does not automatically mean more valuable for training. A trajectory may contain irrelevant steps or omit critical decisions. Performance on a benchmark does not establish reliability across every real business workflow.
Questions to ask
- Which earlier decisions constrain later actions?
- Can errors and recovery be observed in the records?
- What independent check establishes completion without unwanted side effects?