Solving a hard problem once is different from carrying a long piece of work through new evidence, interruptions and the consequences of earlier actions. We call the second ability durability, and this paper argues that useful capability needs it as much as it needs intelligence.
To measure what durability adds on its own, we held the base model fixed. We graded 6,000 runs blind on GPT-5.5, Opus 4.6 and Gemini 3 Pro across five professional workloads, and followed 384 work programmes of up to 42 days.
What we found
- Task fidelity on day 30 reached 95.2%, against 63.9% for the strongest alternative we tested, a leading agent framework.
- Runs with no violations rose from 18.4% to 97.6%, and outputs with unsupported facts or reasoning fell from 8.2% to 0.7%.
- In the 42-day programmes, agents ran a median 109.6 hours before a person had to rebuild their work, up from 7.4. Accepted work rose from 16.2% to 24.3%, and the median human time each programme needed fell from 41.6 hours to 13.7.
- The same programmes used 82% fewer tokens: 2.5 billion per 42 days, against 13.9 billion.
- Keeping information is not enough. A system that kept 94% of its information but did not revise the decisions that depended on it made 22 violations per thousand decisions; the full system made 6.
- The same continuity lets a system build up experience and switch to a stronger model in the middle of a project, carrying its tested methods and their evidence across.