What coding agents actually read is not the technical docs — across 557 sessions and 3,033 documentation interactions, six in ten involve instruction files and working notes

1The readership changed; the writing did not

Technical documentation has been written for human developers. Yet the share of software changes actually authored by autonomous coding agents keeps rising.

Which documents do agents consult, when, and what happens next? This study answers from records rather than guesses, drawing on two public datasets: 557 agentic coding sessions (94,813 development events, of which 3,033 are documentation interactions) and 33,097 agentic pull requests (690,260 file-level change records).

2What is actually being read

Six in ten documentation interactions go not to documents written for a human reader but to the agent's own instruction files and working notes. Interactions with the API references teams have laboured over come to 1.3%. It is a figure that unsettles the premise of polishing technical documentation so that agents will read it.

3The read-then-write sequence is not observed

The link between consultation and editing also defies intuition. The probability that an edit immediately follows a consultation is 0.002, and the unadjusted three-event lift is 1.05. A stage-adjusted model puts it above unity, but this is far from a clean causal sequence. No explicit documentation-based validation sequence was observed at all, and consultation was associated with less immediate testing rather than more.

The trigger for consultation is mostly self-initiated (70.2%) rather than failure-driven (7.5%). Among multi-commit pull requests that change both code and documentation, code is touched first 4.7 times more often. Documentation trails code.

4Questioning the premise of agent-friendly documentation

From these traces the authors derive a descriptive model of agent-documentation interaction as a two-lobed cycle rather than a linear journey, and show that two widely assumed properties of agent-friendly documentation — actionability and verifiability — lack consistent behavioural support. Of the 1,900 arXiv papers this site holds as of 2026-09-02, 37 have cs.SE as their primary category.

Amid a field dominated by papers proposing models and methods, work that reads large volumes of real activity logs to test an existing premise stands out. This article is our own summary and does not warrant the correctness of the claims.