From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation
Ranking
Overall
82
Content
100
Popularity
41
Observed public metrics from 1 member.
Merged summary
TL;DR - An empirical study of more than 33,000 agentic pull requests and 557 coding-agent sessions finds that agents primarily use instruction files and working notes rather than conventional technical documentation. The results challenge assumptions that agent-friendly documentation reliably prompts action or validation.
- Agent-facing artifacts represented 60.5% of documentation interactions, versus 10.6% for classical technical documentation and 1.3% for API references.
- Evidence that documentation consultation leads to code editing was mixed: the adjacent transition probability was 0.002, while a stage-adjusted model found elevated odds (OR 1.33).
- No explicit documentation-based validation sequence was observed; consultation correlated with less immediate testing (adjusted OR 0.39).
- Consultation was usually self-initiated (70.2%), and code was touched before documentation 4.7 times more often in multi-commit pull requests changing both.
Sources (1)
From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - An empirical study of more than 33,000 agentic pull requests and 557 coding-agent sessions finds that agents primarily use instruction files and working notes rather than conventional technical documentation. The results challenge assumptions that agent-friendly documentation reliably prompts action or validation.
- Agent-facing artifacts represented 60.5% of documentation interactions, versus 10.6% for classical technical documentation and 1.3% for API references.
- Evidence that documentation consultation leads to code editing was mixed: the adjacent transition probability was 0.002, while a stage-adjusted model found elevated odds (OR 1.33).
- No explicit documentation-based validation sequence was observed; consultation correlated with less immediate testing (adjusted OR 0.39).
- Consultation was usually self-initiated (70.2%), and code was touched before documentation 4.7 times more often in multi-commit pull requests changing both.