TLDR Dev 2026-10-08

In August, agents made 257 million web requests versus 131 million human page loads. Docs are only one surface. Agents also answer from help centers, changelogs, community forums, and support archives. In our survey of 329 professionals, 94% said their company has knowledge beyond docs, and just 7% have made all of it readable by agents. The 2026 State of Knowledge Report covers our 2,400-test navigation benchmark, how Anthropic runs Claude Code's docs, and what this means for how you maintain knowledge.Read the report

Streaming tokens as event logs makes refreshed or late-joining clients rebuild partial responses, while rewriting one record for every token overloads the database. The article compares four storage patterns and argues that message appends can return the full response so far, continue live delivery, and save only the completed message to application storage.

Cortex replaced a Lucene-based search architecture with PostgreSQL full-text search, a custom query parser, weighted ranking, and CDC-fed indexes. The change cut public API p95 latency from 13.3 seconds to 1.54 seconds, reduced total memory use by about 40%, and made most edits searchable within seconds.

Software blogs lose readers when they meander, assume too much background knowledge, or outsource key explanations to links. The guide recommends stating the payoff in the title and opening sentences, writing for a specific reader, and keeping external references optional.

Chrome 155 will ship a memory-safe JPEG XL decoder built in Rust, reversing the browser's earlier decision to remove support. Chrome's reach could make the format mainstream, especially for high-fidelity photography, lossless compression, progressive decoding, and lossless JPEG transcoding.

The essay examines five objections to AI coding: code as art, job displacement, dehumanizing work, rule-based views of software engineering, and support for powerful AI companies. It argues that several critiques depend on debatable assumptions about what software work is for, while locating the strongest human concern in labor relations and control of production.

Is your existing CMS holding you back? Drupal delivers an enterprise-grade, open-source foundation designed to power high-traffic, secure and future-proof digital experiences. Test the AI integrations, the decoupled flexibility, and see why 400k+ global sites trust it - 100% risk-free. Explore the free trial.

GPT-6 is rolling out in ChatGPT with Intelligent UI, which can combine text, graphics, forms, charts, and interactive tools in a response. Paid tiers use GPT-6 Sol, Free and Go use GPT-6 Luna, and the release can begin responding while it continues reasoning or searching.

Claude Haiku 5.5 targets high-volume, latency-sensitive work such as summarization, classification, browser use, and subagent tasks. Anthropic says average running costs are about 75% lower than Haiku 4.5, with adjustable effort and availability across its platform, AWS, Google Cloud, and Azure.

Docker Agent is a CLI plugin for defining and running AI agents from declarative YAML. It supports multi-agent orchestration, MCP and built-in tools, multiple model providers, retrieval, and packaging agents to OCI registries.

Agentic development changes the ideal linter because automated checks run on nearly every coding turn and agents can act on rules that humans might find too noisy. Ultracite pairs fast tools such as Oxlint and Oxfmt with strict rules and agent instructions, while keeping judgment-heavy questions for review.

A remote Upstash Box gives Grok Bot a shell, persistent filesystem, GitHub integration, public previews, schedules, snapshots, and optional browser automation through MCP. The examples show it building a live app, cloning and fixing repositories, and turning a Slack report into a Linear ticket and pull request from a phone.

Vanguard's password reset form silently truncated pasted passwords at 20 characters, while its login form accepted the full value and then rejected it as incorrect. The incident shows why password length limits should be validated visibly in JavaScript or on the server instead of using an HTML maxlength attribute that changes the submitted credential.

TinyFish is the all in one Webstack for agents. Give them the tools to search, extract, browse, and act on the web on one platform. Now claim your 30% bonus

A Firestore-to-SQL migration only became faster after the read model moved aggregation to the server, cutting a ten-year view from 56 seconds to 3.7 seconds and shrinking the shared search payload from 18.27 MB to 1.63 MB.

REA connects coding agents to local tools for reverse engineering websites, JavaScript and Electron apps, .NET assemblies, and native binaries.

This visual notation distinguishes React's parent and owner trees so component structure and creation relationships can be read as diagrams instead of inferred from scattered code.

Size Limit enforces JavaScript performance budgets in CI, including bundle size, dependency cost, and estimated download and execution time.