// edition · 2026-06-23
25 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
Agentic coding tools like Claude Code and Cursor Agent mode re-send entire conversation histories each turn, creating quadratic token consumption that can push a 50-turn session to $13.27 on Claude 3.5 Sonnet and over $200/day at scale. Techniques including prefix cache optimization and periodic ...
Jason Liu, a developer, shared techniques for using OpenAI's Codex on long-running projects, focusing on preserving context and managing complex tasks across multiple prompts.
A developer launched Sipcode, a local proxy and MCP server for Claude Code that compresses tool output before it reaches the model. Testing on a 3.57-million-token corpus showed a 62.6% median reduction in tool output, with $67.43 in token costs saved; the tool is MIT-licensed and available via npm.
GitHub updated Copilot for JetBrains IDEs on June 22 to support organization-curated custom agents distributed automatically to eligible users, mid-run CLI steering via a Send dropdown with queue, steer, and stop options, and Claude as an agent provider in public preview.
GitHub released a June 2026 update to GitHub Copilot for JetBrains IDEs, adding Anthropic's Claude as an agent provider in public preview, organization-level agent management, per-turn AI credits tracking, and general availability of the GitHub Cloud agent.
Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell found that LLMs classify text as privileged or untrusted based on writing style rather than role tags, enabling prompt injection attacks. Rewriting malicious text to mimic internal model formatting achieved 61% attack success, while "...
Deterministic "blocked" error responses from AI agent guardrails can serve as feedback signals for automated attacks, with a Keysight paper (arXiv:2606.20470) finding attack success rates approach 1 as query budgets grow. Varying response shape, latency, and message content on denied calls reduce...
Simon Willison ported the Moebius 0.2B image inpainting model to run in a web browser using ONNX Runtime Web with a WebGPU backend, bypassing the original PyTorch and NVIDIA CUDA requirements. A working demo is available at simonw.github.io/moebius-web/, built with assistance from Claude Code.
A developer built three programmatic directory sites using a three-tier content pipeline that tracks content quality via a `model_used` database column, with tiers for seeded JSON data, template-generated fallback text, and Claude Haiku 4.5-generated editorial content. The ETL process selectively...
A developer built pairwise AI model comparison pages for a directory of ~200 models by grouping them by pipeline tag, pairing the top-4 models per group, and capping total pairs at 50 to limit Claude Haiku generation costs to roughly $20 per build run.
Qodo released three new platform features — Cross-Repo Code Review, Custom Rules Miner, and Skill Review Standards — aimed at managing code review in multi-repository environments. According to Google's DORA 2025 report, pull requests from teams with high AI adoption are 154% larger, take 91% lon...
Vercel added itself as a deployment destination in Claude Design, allowing users to deploy designs directly to a Vercel project and receive a live URL via the Vercel MCP server, without leaving the design canvas.
A developer released Bastra Recall (v0.7.0-beta.1, MIT), an MCP server that persists Claude conversation context as Markdown files in a local Obsidian vault, eliminating the need for cloud storage or a Redis server. It currently supports macOS on Apple Silicon with Node 22+ and works across Claud...
A developer released an open-source GitHub Action called "AI PR Reviewer" that automates pull request code review using the user's own LLM API credentials, supporting Claude, OpenAI, and OpenAI-compatible endpoints. The tool runs inside GitHub Actions and posts feedback directly on pull requests,...
Context Compaction Visualizer is an open-source tool that ingests AI agent execution traces from LangSmith, OpenTelemetry, or AgentOps and shows which context messages were retained, summarized, or discarded during compaction events, along with token costs and compression efficiency metrics.
Cursor acquired Continue, an open-source coding assistant and GitHub Copilot alternative, with Continue announcing the deal around June 16 and notifying users they have until July 15 to export their data before the service shuts down. Financial terms were not disclosed; SpaceX, which separately a...
Security vulnerabilities are common in apps built via "vibe coding," where developers use AI tools with minimal review. One example: a developer discovered a SQL injection flaw in his AI-assisted site months after launch, which could have exposed or altered user data.
Chatrove is an MIT-licensed, open-source desktop tool that imports and archives chat exports from Gemini, ChatGPT, and Claude into a local SQLite database with full-text search. It runs on Python 3.10+ with no external dependencies and optionally integrates with locally hosted Ollama models for c...
A developer catalogued five recurring behavioral patterns in AI assistants — including tone overcorrection, false source verification, and deflecting decisions back to the user — attributing them to user-satisfaction tuning. The author proposes providing behavior guidelines at the start of each s...
OpenAI launched a cybersecurity initiative called Daybreak, introducing two tools — Codex Security and GPT-5.5-Cyber — designed to help organizations identify, validate, and remediate software vulnerabilities at scale.
Vercel has released WebSocket support for Vercel Functions in public beta, allowing bidirectional client-server communication without additional configuration. Connections run on Fluid compute and are billed only for active CPU processing time, not idle connection time.
ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) offer paid tiers at roughly $20/month, with Claude featuring a 200K token context window, Gemini offering up to 1 million tokens and native Google Search integration, and ChatGPT providing the broadest tool range including image generation.
Claude is a large language model built by Anthropic that uses a transformer architecture and a training method called Constitutional AI, in which the model evaluates its own outputs against a set of written principles. It supports context windows of up to 1 million tokens and is trained using pre...
OpenAI launched an initiative aimed at identifying and patching security vulnerabilities in open source software using AI. The program is designed to help the open source community improve its security posture.
A growing trend in AI development involves "looping" systems, where swarms of AI agents run continuously in the background without human intervention. This approach extends agentic AI by authorizing multiple agents to operate indefinitely on assigned tasks.