☕️ A Claude Code skill wasted 200K tokens

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🐛 A Claude Code skill wasted 200K tokens

🤖 Block launches open-source AI agent Berd

📌 AWS Bedrock adds multi-agent compute

🔒 Cloudflare adds MCP write controls

🏭 Warp launches Warp Factories

⚛️ React Router v8 ships ESM-only builds

Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🐛 A Claude Code skill wasted 200K tokens LINK

Anthropic cut its built-in /claude-api skill in Claude Code from over 200,000 tokens down to roughly 25,000, after the tool was loading its entire reference library before answering even a one-line question.

The v2.1.234 fix, shipped Monday, now pulls the skill's reference docs on demand instead of inlining everything, an 85%+ reduction that frees context for the repository and the actual work.

The old behavior bundled about 812 KB of material and, when it couldn't detect a project language, loaded docs for C#, cURL, Go, Java, PHP, Python, Ruby and TypeScript plus 26 Markdown files, though only one ~33 KB file was needed.

🤖 Block launches open-source AI agent Berd LINK

Block open-sourced Berd, a desktop application that gives users one consistent environment for running AI agents across different models, tools and projects, now available under Apache 2.0 with builds for macOS, Windows and Linux.

Built with Tauri 2 and React 19, Berd is an orchestration layer rather than a model or runtime, communicating with Block's Goose framework via the Agent Client Protocol while handling projects, sessions, context and configuration on top.

Berd stores conversation history locally and keeps credentials in the OS keychain, ships telemetry disabled by default, and gives each configured agent a distinct role, skills and animated visual identity so users can tell them apart.

📌 AWS Bedrock adds multi-agent compute LINK

AWS added runtime instances to Bedrock AgentCore, a second compute option that runs agents on managed EC2 in the customer's account, keeping existing AgentCore APIs, identity and observability while lifting the previous eight-hour microVM session ceiling.

Instances give agents persistent EC2-backed sessions lasting up to 14 days, with shared file systems, GPU instance types, and Python or container images, and let multiple agents co-locate on one host to collaborate through a shared directory instead of API handoffs.

A new capacity provider primitive defines allowed instance families, OS, networking and storage while AWS handles provisioning, patching and scaling; instances bill at standard EC2 rates plus a management fee, breaking even around 24% sustained CPU utilisation versus microVMs.

🔒 Cloudflare adds MCP write controls LINK

Cloudflare launched WriteGuard, now in private beta, giving MCP servers fine-grained security controls that govern which tools AI agents may use to modify data or trigger actions rather than just read information.

Sitting behind Cloudflare's MCP portal, it intercepts every request, loads the targeted tool's policy, and either passes or blocks it, routing denied and later-failed requests to an auditing service without touching the servers.

Each tool gets a risk tier, with production deploys, merge completions, and bulk deletes flagged critical, and rather than separate agent accounts, it uses existing OAuth credentials plus client and session context for audit logs.

🏭 Warp launches Warp Factories LINK

Warp has launched Warp Factories, an infrastructure layer that lets companies deploy, steer, and manage fleets of coding agents across the standard software development stages without building the underlying tooling themselves.

The system automates any of the triage, specification, implementation, review, and verification phases, works with models like Codex and Claude Code, and integrates with Linear, Jira, Slack, and Teams to fit existing workflows.

Managers get dashboards to compare agent configurations, track token spend, and run self-improvement loops; Warp's own team automates roughly 30-35% of weekly tasks, with humans still required for the rest.

⚛️ React Router v8 ships ESM-only builds LINK

React Router v8 shipped on June 17, 2026 as an ESM-only build with a small set of breaking changes, enabling middleware by default and marking React Router v6 and Remix v2 as End of Life.

The release raises baselines to Node 22.22.0+, React 19.2.7+, and Vite 7+, updates tsconfig and fields to ES2022, and graduates four future flags into standard behaviour that teams could already adopt in v7.

Middleware as a baseline pushes teams to centralize auth, logging, and headers, and migration involves updating peer dependencies, removing deprecated APIs, and switching DOM-specific imports for a package pulling over 50 million weekly npm downloads.

Other news you might like

• What happens to your indexed data when Mistral flips the switch?LINK

• PHP 7.4 To PHP 8.6 Benchmarks, PHP 8.6 JIT PerformanceLINK

• Netflix Open-Sources Agentic Workflow for Causal InferenceLINK

• Introducing LangSmith Tuned EvaluatorsLINK

• Dev uses Claude AI to create native macOS driver for 'obscure' Windows-only printer — Linux container hack enables system-wide Cmd-P printing, driver now available on GithubLINK

• GitHub programmer ports playable Doom to DSLR camera with 3-inch TFT LCD display — Canon EOS 550D with open-source Magic Lantern firmware uses camera's button as controls, even plays soundLINK

• Teleport puts developer Linux desktops under production access controlsLINK

🧰 Trending tools

AnySearch: a search API for AI agents that queries multiple trusted sources in parallel, returning filtered, de-duplicated results as structured data.LINK

Prelint: checks AI-generated pull requests against your ADRs, docs, and prior decisions to catch product drift before it ships.LINK

SKI: enables voice-based coding with Claude Code, Codex, and other AI agents, giving spoken replies so you work hands-free while staying focused.LINK

Dograh: an open-source voice agent platform with a visual flow builder, 30+ model integrations, local model support, telephony, human transfer, and self-hosted QA monitoring.LINK

Zro: routes coding requests to open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through one endpoint without storing your data.LINK

AgentSky: runs managed AI coding agents with auto-recovery, letting you control them from WhatsApp, Telegram, Slack, web, API, or CLI.LINK

📚 Trending papers & reports

Microservice incident repair uses live system monitoring data, not bug reports, to automatically pinpoint and patch faulty code and deployment settings, beating rival tools on cost-effectiveness across a 575-case benchmark.LINK

Automated user interviews use two coordinated AI agents to dig into software feedback, keeping conversations on track and turning ~81% of vague complaints into concrete, actionable product requirements without human interviewers.LINK

Spec-driven test generation makes coding agents first write down a program's expected rules before creating tests, catching ~10 points more bugs at Google and beating human-written tests in over half of cases.LINK

Self-referential test oracles can silently pass buggy software when a fault shifts both the measurement and the expected answer together, and fixing one such test on a live air traffic simulator caught 8 of 46 missed defects.LINK

Security warning triage automatically tests small isolated chunks of code around each flagged issue, confirming ~55% of real vulnerabilities in benchmarks while sparing developers days of manual inspection and full-project fuzzing.LINK

See you tomorrow for a new dose of ☕️ Devshot!