Writing: AI agents generating documentation

Data on how people use agents to generate and maintain documentation.

How AI agents generate and maintain documentation.

The writing layer has three distinct session types, each with different trigger patterns, human involvement, and output quality. 73,058 requests across 3,710 sessions on a single Tuesday in April 2026.

Scale

Metric	Value
Total requests	73,058
Unique sessions	3,710
Requests per session	19.7

The 19.7 requests per session reflects the agentic nature of writing: read existing docs, check the codebase, draft, verify, revise.

By product

Writing requests break down across four products with different levels of human involvement.

Product	Requests	Share	Trigger	Human input
Workflows	35,581	49%	System (webhook/cron)	None
Atlas	34,177	47%	Human (enters URL)	None after initiation
Agent sessions	2,200	3%	Human (Linear issue)	Continuous
Agent sessions	1,100	2%	Human (Slack command)	Continuous

Agentic reading versus writing

Metric	Reading	Writing
Requests per session	3.3	19.7
Average prompt length	47 words	556 words
Average response time	3.7 seconds	8.1 seconds
Median response time	2.9 seconds	4.1 seconds
Average prompt tokens	Not included	4,335
Average completion tokens	Not included	437

Three session types

Directed

Human initiates on a Linear ticket or Slack command. Continuous human involvement. Includes edits, content creation, questions, PR work.

Automated

System-triggered by a webhook (push) or cron job. No human in the loop after setup. The majority of session volume.

Complex

Multi-agent pipelines with large system prompts. High coordination overhead. Atlas (repo-to-docs) is the primary example.

Directed sessions

Humans trigger these via Linear ticket or Slack command and stay involved throughout.

What people ask for

Task type	Share
Edits	27%
Create new content	18%
Questions	15%
PR-related	12%
Discussion	11%
Review	8%
Information architecture	5%
Clarification	4%

PR merge rate over time

AI-generated PRs from directed sessions have improved substantially:

Month	Merge rate
January	47%
February	39%
March	52%
April	74%

Prompt length

Directed session prompts average 1,133 characters. Detailed enough to give an agent direction, short enough that humans write them.

Automated sessions

System-triggered, no human required. The bulk of automated sessions run on a schedule or fire on a code push.

Trigger type breakdown

Trigger	Share
Cron jobs	70%
Push-triggered	30%

Push-triggered sessions produce 2.1 times more merged PRs on average. Likely because they’re responding to code changes and a documentation need is immediate and specific.

Prompt length

Automated session prompts average 1,146 characters. Very similar to directed sessions.

Complex sessions

Atlas (Mintlify’s repo-to-docs generator) is a complex writing pipeline. A single Atlas run involves four distinct prompts.

Prompt type	Prompt size
Pre-analysis	7,000 characters
Coordinator	19,000 characters
Writing agent	5,000 characters
Review agent	3,600 characters

The coordinator prompt alone (19,000 characters) is longer than most documentation pages. You can’t just say “write good” to an agent.

Benchmark

Agentic PR merge rate: 55%+

More than 55% of AI-generated PRs are merged.

How to measure

The most universal proxy is PR merge rate. The share of AI-generated pull requests that a human accepts and merges. Other signals worth tracking:

• Time from generation to merge
• Review comments per PR
• Revert rate

Don’t replace your deterministic tools. AI writing works alongside linters, validators, and style checkers.