A docs-quality agent treats documentation as an executable spec: run the guides in a secure sandbox, follow the steps a developer would, detect drift against reality, and open Linear issues plus GitHub PRs when the docs lie. That is the high-level pattern — not a dump of internal Coinbase IP.
The problem
Developer documentation drifts. APIs change, screenshots rot, copy-paste commands fail. Humans catch some of it in review; most of it ships as quiet debt until a customer or an on-call engineer hits it.
Why an agent
Static linters catch broken links. They do not walk a multi-step guide the way a developer does. An agent with tools and a sandbox can attempt the path and report where the path and the world diverge.
Pattern (teaser)
- Ingest the docs surface you care about.
- Execute steps in a secure sandbox (no production credentials; controlled network).
- Compare outcomes to expected success criteria in the guide.
- Act — raise a Linear issue and, where safe, open a GitHub PR with a proposed fix or a failing repro.
Built on the Claude Agent SDK for tool use and long-running agent loops. Complements (does not replace) LangGraph systems elsewhere in the stack — same production bar: observability, boundaries, human review on high-impact writes.
What this is not
This post does not cover internal CDP architecture, data stores, or proprietary prompts. Those stay internal. The transferable idea is docs-as-spec + sandbox execution + issue/PR loop.
When to use it
- High-churn developer platforms
- Guides that must stay runnable
- Teams that already trust agents for review-shaped work with human approval
If you are wiring agents into docs, tools, or MCP surfaces and the hard part is production boundaries — not the demo — that is the class of problem these field notes are for.