Empirical verification of DOX (AGENTS.md) impact: +193% session turn capacity in developer logs

[Korck-lab]

Hello DOX team!

We wanted to share some exciting empirical data backing the design philosophy of the DOX (AGENTS.md) framework.

We developed claude-code-optimizer to ingest and analyze developer sessions. During a deep audit of 9,911 real-world Claude Code session logs, we isolated the metrics of sessions utilizing custom agent rulesets (AGENTS.md / DOX) against those without.

Here is what the data showed:

Metric	With DOX (AGENTS.md)	Without DOX	Difference / Impact
Average Human Turns / Session	4.87	1.66	+193% turn capacity
Average Prompt Input Tokens	73,097.8	16,537.6	Deeper code context navigation
Average Tool Errors / Session	0.68	0.37	Contextually aligned tool executions
Average API Cost per Session	$10.48	$2.09	Deeper execution per cached session

Key Takeaway

Sessions using DOX are almost 3x longer on average (measured in interactive human turns) and navigate significantly larger contexts. This suggests that structured instruction hierarchies successfully align the agent, allowing developers to execute longer, more complex coding tasks without the agent getting lost or drifting off-task.

Thank you for creating DOX! We hope this quantitative validation is helpful to the community.

[BernhardBaumrock]

@Korck-lab I don't understand your issue, maybe you can elaborate a bit more for people not so deep into the topic?

Some numbers seem to be higher with DOX where I would assume that lower numbers are better? What I find especially confusing is average errors with DOC 0.68 whereas without shows 0.37 which would be a lot better? Less errors is good in my world?! :)

But your takeaway seems to be very positive. I'm confused (and just discovered DOX today)

[rahmed-dev]

Following