TLDR Dev 2026-06-18

Same input. Same prompt. Different output. That's the reality of testing AI agents that write code, and most teams are shipping without solving it.Nick Nisi from WorkOS tackled this by building eval systems for two AI tools:- npx workos@latest, a CLI agent that installs AuthKit into your project- WorkOS agent skills that power LLM responses about SSO, directory sync, and RBAC.The post covers how to test against real project structures, score output that's different every time, and catch when your agent makes up methods that don't exist.Learn more about evals →

Browser Use's cloud browsers have been optimized to start quickly, remain isolated, and maintain low costs, now operating at $0.02 per hour with sessions starting in under a second. This was done by using Firecracker for lightweight virtual machines while implementing a control plane for dynamic scaling.

Dirge is a highly efficient coding harness developed in Rust that improves budget model performance while still having a minimal memory footprint. Unlike traditional harnesses that focus solely on model capabilities, Dirge's supportive framework, with features like error correction, session continuity, and a memory management system help optimize model quality across different contexts.

A clear division of labor exists in agentic coding where humans make the majority of planning decisions while the AI handles technical execution. Success in these tasks depends more on the user's domain expertise than their formal programming skills, allowing professionals in various non-technical fields to get results comparable to those of software engineers. Recent trends show a shift toward complex, end-to-end agentic work and a measurable increase in the estimated economic value of the tasks being completed.

Yak shaving refers to the process of engaging in a series of tangential tasks that divert from the original goal, showing how building projects from scratch can lead to unexpected challenges, discoveries, and joy.

Conversations with others often lead to greater problem-solving and understanding than solitary thinking can provide as they introduce a feedback loop that refines thoughts in real-time.

The infra. Token metering, model routing, usage variance. Vercel, Replit, Glean, and Supabase use Orb to handle it. See how 80 AI companies actually price in the 2026 State of AI agent pricing report.

Lore is an open source version control system designed for scalability, particularly for projects with large binary assets like games. It has a centralized, content-addressed architecture that allows for efficient data storage and retrieval, with an emphasis on immutable revisions and low-overhead branching.

eve is an open-source agent framework that makes building, running, and scaling agents easier by providing built-in production capabilities like durable execution, sandboxed computing, and human-in-the-loop approvals. With a clear structure, agents can be defined through a directory of files that dictate their components, actions, and connections.

TREX is a code review tool that integrates code execution with AI-driven testing to identify bugs that may not be apparent in static analysis alone. It uses a multi-agent architecture, where a Greptile reviewer orchestrates specialized TREX agents that run tests in isolated environments, providing artifacts such as screenshots and logs to verify findings.

An analysis of various LLMs in triaging security vulnerabilities showed that a council of four LLMs achieved an impressive 86.2% unanimous agreement on findings, while lower reasoning efforts consistently performed poorly. Although higher reasoning generally correlated with better results, there were exceptions where models struggled with complex tasks.

Image compression techniques, including lossy and lossless methods, exploit characteristics of human visual perception to reduce file sizes. Formats like JPEG use a combination of chroma subsampling, discrete cosine transforms, and quantization for compression without noticeable quality loss, while GIF uses indexed color palettes and Lempel-Ziv-Welch algorithms to store limited color data.

Foxit's open-source MCP server gives AI agents access to PDF conversion, OCR, merge, and signing. Works with Claude Desktop, Cursor, and GitHub Copilot. Set it up in under 15 minutes.

Gzip can be used as a language model through compression techniques that predict and generate text continuations by scoring potential outputs based on how well they compress with a given context.

mac-ocr is a macOS command-line tool that uses Apple's Vision framework locally to recognize text from images and PDFs and create searchable PDFs.

Exa Agent is a new API that combines frontier language models with optimized web search tools to provide high-performance research, list-building, and data enrichment capabilities.

Anthropic has released a playbook of practical exercises and agentic workflows to navigate every stage of the business lifecycle for AI-native startups.