The Harness and Memory Architecture Behind Self-Updating Docs

Mintlify rebuilt its self-updating documentation system around background and runner agents that share a git-backed memory repository, respond to triggers such as code merges or Jira events, and typically produce documentation pull requests for human review. The background agent continuously evaluates PR outcomes, integration data, analytics, and run logs to modify triggers, skills, and memory, with the goal of maximizing merged documentation PRs while maintaining an acceptance rate above 80%; internally, this approach reportedly increased acceptance from 70% to 90%. Rather than relying on third-party agent harnesses, the company built its own using Vercel’s AI SDK and custom tools for subagents, session management, skills, and sandboxed shell commands, achieving comparable benchmark performance to Pi with lower token usage. A central design choice is fixed-schema JSON memory for releases, users, style guides, settings, and integrations, which provides reliable context, enables offline evaluation, and reduces the ambiguity of free-form agent memory. The architecture also supports over-the-air updates by instructing background agents to adapt customer-specific setups when schemas, skills, or products change, while JSON run logs and forked session trees preserve agent histories for analysis and improvement.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.