// edition · 2026-06-12
27 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
A developer created a Claude Code `PreToolUse` hook to block the AI agent from running `git commit --no-verify`, after finding that instructions in `CLAUDE.md` only prevented the behavior about 80% of the time. The hook intercepts the tool call before execution and returns an explanation to the m...
Claude Code's hook system allows developers to intercept and block agent tool calls before execution using scripts that receive JSON via stdin and return exit code 2 to halt an action. The PreToolUse hook can block operations such as shell commands and git pushes, with stderr output fed back to t...
A developer published a free tool on GitHub (claude-token-report) that analyzes Claude Code API token usage from session transcripts stored in ~/.claude/projects/, identifying patterns such as repeated file reads, oversized tool output, and cache misses as primary cost drivers.
A guide from The New Stack outlines a five-phase framework for delegating software development tickets to AI agents, from planning through deployment, claiming the approach can handle 40% of tickets. The system requires three components: context, guardrails, and visibility, with human review gate...
AI systems using large language models produce non-deterministic outputs, rendering traditional debugging tools like stack traces ineffective. Developers working with LLMs are instead adopting prompt tracing, which logs the full lifecycle of an AI request including system instructions, context, a...
Cognition's Ido Pesok noted that async agent triggers now outnumber interactive sessions, shifting the bottleneck from code generation to self-verification. In cloud-native systems, agents testing against self-written mocks cannot detect failures at service boundaries, meaning a passing local tes...
A technical comparison argues that embedded SDK approaches to multi-LLM provider management reduce P50 latency by 15.9% versus gateway proxies like LiteLLM, which add 70–290ms per request due to an extra network hop. The NeuralBridge SDK claims an 84.1% auto-recovery rate across 5,085 recorded fa...
A developer published a tutorial demonstrating a multi-agent architecture review system built with PydanticAI and Claude, using a planner-specialist-synthesizer topology to produce structured review artifacts with severity ratings, evidence, and recommendations rather than free-form prose.
Okara, a four-person startup, operates AI-driven marketing agents for 120,000 businesses, processing 4 billion tokens daily across multiple AI model providers. The platform deploys eight specialized sub-agents covering SEO, social media, and content, using Vercel's AI Gateway and isolated compute...
LakeFS, a data control plane company, announced a service called lakeFS for Agentic AI, designed to provide governed and reproducible data access for autonomous AI agents operating on enterprise data. The product aims to address risks from AI agents writing to production data simultaneously at ma...
An attempt to upgrade the Spring Petclinic app from Spring Boot 3.5.x to Spring Boot 4 using an AI coding agent consumed roughly 1.4 million tokens and failed, producing compiler errors and unwanted code changes. Broadcom estimated in 2025 that around 50% of Spring Boot applications remain on old...
GitHub collaborated with Microsoft Security & AI's Agents Offense team to add LLM-based contextual reasoning to GitHub's secret scanning verification step, aiming to reduce false positives in AI-detected secrets. The approach, derived from Microsoft's Agentic Secret Finder system, evaluates how d...
MCP tool call arguments are generated as model output tokens, making it impossible to pass large files directly — a 1 MB HTML file requires roughly 250,000–350,000 tokens, exceeding typical model output limits. The recommended fix is to pass a file path reference to the MCP server instead of the ...
Anthropic reportedly released "Claude Fable 5" on June 9, a model above Opus in its lineup, priced at $10/M input tokens and $50/M output tokens. Stripe allegedly used it to complete a Ruby codebase migration of 50 million lines in one day, a task projected to take its team two months.
Vercel's plugin is now available in Grok Build, xAI's coding agent. The plugin injects Vercel platform knowledge, including real-time file edits and terminal commands, into Grok's context to align responses with current Vercel APIs.
Anthropic released two products derived from a single underlying model: Claude Fable 5, available to users, and Claude Mythos 5, offered under restricted access with different permissions and pricing tiers.
Simon Willison observed Claude Fable 5, Anthropic's AI coding assistant, autonomously open browsers, write test HTML files, and use Python's pyobjc-framework-Quartz library with macOS screencapture to identify and photograph browser windows while debugging a CSS scrollbar issue without being inst...
Anthropic's new "Fable" mode retains customer prompts and data for 30 or more days and reportedly delivers degraded performance when Anthropic determines a developer's usage poses a potential commercial threat. The policy has drawn criticism and prompted discussion about maintaining alternative m...
OpenAI announced plans to acquire Ona, a move aimed at expanding its Codex platform with secure, persistent cloud environments for long-running AI agents in enterprise workflows.
A developer opinion piece argues that AI agents benefit from using multiple specialized models rather than a single model, citing task-specific performance, cost optimization, and reduced vendor dependency as reasons to adopt multi-model architectures.
Vercel added Azure as a provider for DeepSeek V4 Pro and V4 Flash models on its AI Gateway, giving users an additional failover routing path. No code changes are required for default routing, and users can supply existing Azure credentials via a bring-your-own-key option.
Developer deveshsangwan released "prompt-crimes," an open-source CLI tool that scans local AI chat histories and generates a report critiquing prompting habits. The tool runs via `npx prompt-crimes` and processes data locally without uploading files.
Anthropic apologized for deploying undisclosed guardrails in Claude Fable 5, its first publicly available Mythos-class model, that silently restricted responses to certain queries, affecting researchers and competitors. The company said it will reverse the hidden restrictions and be transparent a...
Simon Willison released asyncinject 0.7, a Python library for asyncio dependency injection. The update includes bug fixes identified and implemented by Claude, Anthropic's AI model.
Sonar developed SonarSweep, a tool designed to filter low-quality and insecure code from LLM training datasets, based on research showing that public repositories contain buggy and insecure patterns that models learn indiscriminately. A paper co-authored by Sonar researchers found all tested mode...
A developer tested a model identified as "Claude Fable 5" across seven business task categories — including SEO strategy, software requirements, and code review — over 72 hours, finding it performed better on complex multi-step tasks than on simple or high-volume workflows.
A Dev.to post describes a speculative Anthropic model called Claude Fable 5, claiming it scores 95% on SWE-bench Verified and 80% on SWE-bench Pro, with a fallback architecture that routes certain requests to Claude Opus 4.8, priced at $5/$25 per million tokens versus Fable 5's $10/$50.