State of the art: the best evidence, and the practice most adopted
Step 3. Upper bounds (what experiments achieved) are kept apart from lower bounds (what they found impossible or harmful).
Shape: state of the art
Best evidence
| Question | Holder | Result | Grade |
|---|---|---|---|
| Does a model raise output on writing tasks? | Noy 2023 | 453 professionals; time down 40%, quality up 18% (secondary) | measured |
| On a coding task? | Peng 2023 | 55.8% faster with the assistant, on one task (secondary) | measured |
| On consulting tasks inside the model's reach? | Dell'Acqua 2023 | 758 workers; 12.2% more tasks, 25.1% faster (secondary) | measured |
| Across 106 experiments, is the combination better than either alone? | Vaccaro 2024 | no, on average: see against | measured |
These are upper bounds on one side of a task type. None measures a project of weeks with checks run by a program.
Most adopted
| Practice | Holder | Evidence of adoption |
|---|---|---|
| A file that briefs coding agents | Agentic AI Foundation 2026 | over 60 thousand projects, the site's own count |
| Workflows before agents, with pauses for human feedback | Anthropic 2024 | the owner's account |
Adoption is a descriptive count. No study found measures the effect of either.
Not measured anywhere in this pool
- A person, a model and checks run by a program on one project over time.
- Whether a rule such as "the model decides nothing" changes the outcome.