// edition · 2026-06-18
28 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
A workflow for using Claude Code to refactor large codebases recommends loading full dependency context before making changes, using a CLAUDE.md file to encode rules and progress across sessions, and requiring Claude to summarize a file's purpose before modifying it.
Claude Code CLI v2.1.91 includes more than 40 built-in slash commands grouped into categories covering session configuration, session reflection, session management, and context management. Commands include options for model selection, cost tracking, conversation branching, context compaction, an...
Claude Code stores per-session token usage logs in JSONL format under `~/.claude/projects/`, while OpenAI Codex writes cumulative token count events to `~/.codex/sessions/`. Both log files break down usage into uncached input tokens, cache reads, cache writes, and output tokens.
A comparison of three AI coding assistants — Cursor ($20/mo), GitHub Copilot ($10/mo), and Windsurf ($15/mo) — finds all three support the same underlying models, with differences primarily in UX, IDE integration approach, and pricing.
A developer released "ccteams," an npm package that installs pre-configured sets of Claude Code subagents into software projects via a single command. The package ships with eight agent teams covering stacks including Next.js/TypeScript, Go, Python FastAPI, Ruby on Rails, and a stack-agnostic gen...
Block built BuilderBot, an internal AI coding agent managed via Slack, that runs over 200,000 operations daily and merges roughly 1,500 pull requests per week — about 15% of all production code changes at the company. The tool, built on Block's open-source Goose framework, picks up tickets from L...
GitHub Agentic Workflows use GitHub Actions and AI agents, including Claude and Copilot, to automate repository tasks such as code review, pull request analysis, and CI/CD integration. The system triggers automated actions based on repository events like pushes or pull requests.
A developer built a 14-gate validation harness for Claude/Codex-generated React apps after identifying four failure modes in multi-page output: inconsistent copy, broken routing, shared state drift, and false completion claims. Tested across 10 demos and 54 screens in four business domains, the h...
GitHub added prompt caching and on-demand tool loading to Copilot in VS Code to reduce repeated context sent to models during long agentic sessions. The company also expanded "Auto," a model routing feature that selects which AI model handles a given task based on task complexity and real-time mo...
A developer released Kiro-Ception, an open-source MIT-licensed plugin for the Kiro AI coding assistant that adds persistent memory across sessions, projects, and machines. It runs locally using SQLite and CPU-based embeddings (~80MB), with search latency under 10ms, and optionally supports multi-...
Vercel held its Ship 2026 conference in London with over 2,500 attendees, announcing Vercel Connect for scoped temporary credentials, an open-source agent framework called eve, and a collection of agent-building tools called the Agent Stack.
Vercel removed CLI-specific deployment limits, allowing teams and AI agents to deploy from local machines and external CI/CD pipelines without previously imposed restrictions.
AWS launched a native iOS app for Kiro, its AI-assisted development environment, at the AWS New York Summit. The app lets developers monitor, review diffs, and approve agentic coding sessions remotely, with compute running in AWS's cloud backend across three session modes: Chat, Spec, and Autonom...
AWS announced two preview features for its DevOps Agent at AWS Summit New York: a release readiness review that checks code changes for dependency risks and compliance before merging, and an autonomous release testing capability that generates change-specific test plans for web and API applications.
A developer proposed replacing raw screenshots fed to AI coding agents with structured JSON payloads that explicitly name the target UI element and intended action, reducing token cost from several thousand to roughly 700 per interaction.
A Dev.to author outlined a six-part configuration method for Anthropic's Claude Projects, arguing that effective use requires specific identity and rules blocks in system prompts, uploaded knowledge files, and separate projects for separate workflows rather than a single general-purpose project.
Tailscale updated its Aperture platform with a chat interface, universal data connectors for MCP and APIs, and sandbox support, aimed at managing AI agents in enterprise environments. CEO Avery Pennarun said AI agents require identity management, access controls, and logging to operate safely wit...
Anthropic's Claude Design platform, which connects design systems with code workflows, has received updates including improved design system import, direct editing controls, and tighter integration with Claude Code to allow teams to transfer work between visual design and development.
Chainguard expanded its Agent Skills platform with a public registry of over 1,000 hardened AI agent skills, a private registry, and a hardening-as-a-service tier for internal skills. The system scans for vulnerabilities including over-permissioned scopes and credential harvesting, then uses AI t...
AWS launched AWS Context at its New York Summit, a service that automatically maps relationships across a company's existing data sources into a knowledge graph. The service gives AI agents access to governed data relationships, business rules, and domain knowledge at runtime using open data form...
Anthropic released Claude Opus 4.8 on May 28, 2026, removing the `budget_tokens` parameter (now returns a 400 error) and adding a fast-throughput mode offering up to 2.5× output speed at $10/$50 per million tokens, mid-session system messages, and a reduced prompt-cache floor of 1,024 tokens.
Vercel launched eve, an open-source framework for building AI agents, at its Ship conference in London. The framework structures each agent as a directory of files, uses Vercel's Workflow SDK for durable conversations, and includes built-in integrations with Slack, Discord, Microsoft Teams, GitHu...
Vercel launched Vercel Connect in Public Beta, a service that issues short-lived, scoped tokens for apps and agents to access external services including Slack, GitHub, Salesforce, and Figma without storing long-lived credentials. Pricing starts at 5,000 token requests per month at no cost on the...
Z.ai released GLM 5.2, an open-weights model with 744B total and 40B active parameters, scoring 51 on the Artificial Analysis Intelligence Index, up from 40 in version 5.1. The update expands the context window from 200K to 1M tokens and maintains the same pricing under an MIT license.
Chinese AI lab Z.ai released GLM-5.2, a 753B-parameter open-weights text model under an MIT license on June 16th, with a 1 million token context window. It ranked first on the Artificial Analysis Intelligence Index with a score of 51 and second on the Code Arena WebDev leaderboard, priced at $1.4...
Charity Majors wrote that AI shifted code generation from expensive and time-consuming to effectively free and instant in 2025, turning code from a curated resource into something disposable and regenerable, and argued this requires more engineering discipline, not less.
A software developer argues that AI agents function as an additional interaction layer over existing business software rather than replacing applications, citing continued reliance on databases, APIs, and workflows. The piece notes AI agents perform well on repetitive tasks but remain prone to er...
Claude and ChatGPT differ in key ways for small business automation: Claude handles longer documents and asks clarifying questions before generating code, while ChatGPT produces faster output on shorter, focused tasks but may silently make incorrect assumptions in complex workflows.