AI Agents Now Outpace Humans Reading Docs, But Most Content Isn't Ready

·5 views

Mintlify's 2026 State of Knowledge Report found that AI agent requests to documentation sites reached 257 million in August, nearly double the 131 million human page loads recorded in the same period. Despite this surge, only 7% of surveyed companies had made their knowledge surfaces fully readable by AI agents. A key challenge highlighted is documentation drift, where product changes update some pages but leave outdated information in quickstarts, support articles, and integration guides that AI coding assistants may still retrieve. The article argues that keeping documentation aligned with the product should become a standard part of the software release process, not an afterthought. A tool called Thally is introduced to address this by connecting product repositories to documentation platforms and automatically opening evidence-backed pull requests when content may need updating.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Why Uptime Metrics Fall Short as Reliability Benchmarks for AI Agents

Traditional uptime and HTTP success rates are insufficient measures of reliability for AI agent systems, as an agent can return a healthy status code while failing to complete tasks correctly or safely. Experts recommend splitting agent reliability into distinct indicator families covering task success, safety controls, latency, efficiency, and escalation behavior. Safety-related metrics, such as the rate of authorized versus unauthorized actions, must never be averaged into general availability scores, since a single policy violation can be masked by otherwise healthy aggregate numbers. Business outcomes, like whether a notification was delivered or acted upon, should be tracked separately from runtime completion using a shared decision ID to enable later reconciliation. SLO targets for agents should be grounded in product risk, user expectations, and observed baselines rather than copied from unrelated systems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer automates MikroTik router to bypass LaLiga's Cloudflare IP blocks in Spain

Spanish courts granted LaLiga the right to order ISPs to block Cloudflare IP addresses during football matches to combat pirate streams, but the blocks also take down unrelated legitimate websites sharing those IPs. A developer frustrated by unexplained outages built an automated workaround on a MikroTik router running RouterOS 7. The setup queries a third-party DNS service called Hayahora every five minutes to detect when a blocking event is active, using more than ten DNS responses as the trigger threshold. When blocking is detected, a firewall rule automatically routes only Cloudflare-bound traffic through an existing VPN, leaving all other traffic on the normal ISP connection. The rule disables itself automatically once Hayahora signals that the blocking event has ended.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds 100+ client-side browser tools to cut through ad-heavy utility sites

A developer created FroggoTools, a growing collection of over 100 lightweight, browser-based utilities including calculators, converters, date tools, and generators. The project was motivated by frustration with existing tool sites that load slowly and interrupt users with cookie banners and popups. All tools are designed to run client-side where possible, avoiding server requests to improve speed and protect user privacy. Building a consistent user experience across dozens of tools proved more challenging than expected, requiring significant time on shared design decisions. The developer notes that marketing and distribution have emerged as the hardest part, as a technically sound project still needs visibility to find users.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Misconfigured Validation Rule Wasted 40 Minutes of Dev Work

A developer building a configurable form system lost a full editing session after discovering that a validation rule — 'Required' — had a mandatory sub-type dropdown with no default value. The incomplete rule appeared valid within the form editor but was only flagged as broken at the point of saving, after 15+ questions had already been configured. The core problem was that the system could not distinguish between a rule still being set up and one that had been left incomplete, since both stored identically as a null value. The author argues that such issues can be prevented by making invalid states unrepresentable, setting safe defaults, or surfacing errors immediately when a user leaves a half-configured field. All three approaches are preferable to deferring validation to submission, which creates long, frustrating feedback loops for form authors.

0 comments Read more at DEV Community