AI agents need instructions and checking
Yi-Hung Chou, June Hyung Kim, Vasisht Kartik, Victoria Jackson et al. · arXiv · 24 Oct 2026
Source finding
Eleven developers spent an average of 33.4% of observed session time formulating prompts or checking AI responses: 11.3% and 22.1%, respectively.
Our interpretation
Businesses should plan these activities as work and measure them alongside result quality and rework. This time share is not a productivity loss.
Method and evidence
Workplace shadowing with manual coding of approximately 15 hours, 255 analysed prompts and 4.6 hours of follow-up interviews. Unreviewed preprint; exploratory evidence.
Limitations and interests
Small purposive sample from two organisations, short sessions and possible observer effects. No productivity or business-value estimate. Two authors were contractors at a participating company; recruitment used personal contacts.
What businesses can do
Track instructions and response checking separately from execution. Assess quality, rework and total lead time per completed task.
Where to start
Start with the abstract, Section 4.1 and Figure 2 on time allocation, then Section 3 on recruitment and coding.