// edition · 2026-06-17
31 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
A tutorial describes how to redirect Anthropic's Claude Code CLI tool to a local Ollama server running open-source models such as Qwen 2.5 Coder 7B, bypassing Anthropic's API entirely. The setup involves installing Ollama, pulling a Qwen model, and launching Claude Code with the flag `--api-url h...
Claude Code v2.1.59 introduced an Auto Memory feature that automatically records user corrections and preferences during sessions and reloads them in future sessions. Memory files are stored in the user's home directory under ~/.claude/projects/, are machine-local, and are distinct from the manua...
A developer outlined strategies for reducing LLM costs, distinguishing between caching and routing to lower expected costs versus runtime guardrails to prevent runaway spending from retry loops, long contexts, and uncontrolled tool calls. The post promotes an open-source tool called ai-costguard ...
Anthropic reversed a planned billing change on June 15, 2026, the day it was set to take effect, leaving Agent SDK, claude -p, and third-party app usage drawing from regular subscription limits. The change would have moved programmatic Claude usage to a separate credit system at standard API rate...
Vercel outlined its "Agent Stack," a set of tools for building AI agents comprising AI SDK for model-agnostic calls, AI Gateway for routing across multiple providers, Workflow SDK for durable step-based execution, and Vercel Sandbox for isolated VM environments.
LLM agent token costs stem from two distinct sources: verbose one-off tool outputs (e.g., fetched web pages averaging 2,500 tokens per 10 kB) and repeated standing context such as system prompts and tool schemas resent on every turn. The two problems require separate solutions — content compressi...
Anthropic paused a planned billing change to its Claude Agent SDK on June 15, the day it was set to take effect. The change would have separated Agent SDK usage into capped monthly credits ranging from $20 for Pro subscribers to $200 for Max and Enterprise tiers, billed at full API rates.
Vercel launched "Vercel for Enterprise Apps and Agents," a platform for deploying internal AI agents behind corporate identity providers. It includes Vercel Passport for access control via Okta or Microsoft Entra, Vercel Connect for short-lived credentials to services like Slack and Snowflake, an...
Maxim AI offers Bifrost, an open-source AI gateway that routes Claude Code traffic through a centralized endpoint, enabling enterprises to manage API credentials, track token spend by team, enforce access controls, and aggregate logs across engineering organizations.
Vercel increased the maximum session duration for Vercel Sandboxes from 5 hours to 24 hours, available on Pro and Enterprise plans. The change targets workloads such as large-scale data processing, end-to-end testing pipelines, and long-running agentic workflows.
A developer using Claude Code to add a multiple-program feature to a gym scheduling app reported the tool broke existing functionality twice before successfully implementing the change, after the developer reverted changes and provided more explicit instructions with screenshots.
A developer built a portfolio reviewer called Leon for getfolio that uses a deterministic rules engine to calculate scores across five weighted pillars, restricting the LLM to generating explanatory text only after finding that LLMs consistently produced inflated and inconsistent scores when give...
Judy AI Lab identified three recurring failure patterns in widely shared office AI prompt templates: missing context, unspecified output format, and no fallback when the model lacks information. The analysis covers eight task types including email replies, meeting summaries, and competitive analy...
Anthropic's Claude AI is being used by UI/UX design and development teams to automate tasks such as summarizing user research, generating design documentation, and assisting with code. The article describes Claude's application in design workflows, including analyzing interview transcripts and dr...
AWS moved its FinOps Agent into public preview, a tool that investigates cloud cost anomalies by correlating billing data with CloudTrail records to identify root causes and responsible owners. The agent integrates with Jira and Slack, answers natural-language cost queries, and operates with a pr...
OpenAI introduced Deployment Simulation, a method that uses real conversation data to predict AI model behavior before release, with the goal of improving safety evaluations prior to deployment.
Wiz Security researchers discovered a hardcoded Supabase API key in Moltbook's client JavaScript on January 31, 2026, exposing 1.5 million API tokens and 35,000 email addresses. On February 3, CVE-2026-25253 (CVSS 8.8) was disclosed for the OpenClaw AI agent platform, with 12,812 of 42,000+ publi...
Anthropic's Claude Opus 4.8 scored 69.2% on SWE-bench Pro, outperforming GPT-5.5 at 58.6% and Gemini 3.5 Flash at roughly 54%. The update adds support for spawning approximately 1,000 parallel subagents in Claude Code, claims a 2.5x speed increase, and reduces unreported code flaws by roughly 4x ...
Vercel released eve, an open-source agent framework that includes durable execution, sandboxed compute, human-in-the-loop approvals, and evals built in. The framework, which Vercel uses internally for its own agents including v0, organizes agents as file directories and deploys on Vercel's infras...
Vercel released eve, an open-source framework for building and deploying AI agents, now in public preview. Agents are structured as file directories and include durable execution, sandboxed compute, human-in-the-loop approvals, and subagents, deploying via standard Vercel infrastructure.
Vercel added GLM 5.2 to its AI Gateway, accessible via the model identifier `zai/glm-5.2` in the AI SDK. The model features a 1M token context window, up from 200K in GLM 5.1, and is designed for long-running engineering tasks.
A developer released an open-source API proxy (MIT-licensed, self-hostable) providing access to Claude, Qwen, and DeepSeek models without identity verification, accepting USDT payment, citing Anthropic's reported KYC requirement taking effect July 8.
SpaceX announced it will acquire Cursor, an AI coding platform, for $60 billion, per an SEC filing. The deal, originally agreed to in April with a $10 billion breakup fee alternative, is expected to close in Q3 2026.
SpaceX agreed to acquire AI code editor Cursor for $60 billion in stock, days after its IPO. The deal is intended to bolster SpaceX's AI division; the company cited a $26 trillion addressable market in AI to IPO investors.
SpaceX agreed to acquire Anysphere, Inc., maker of AI coding tool Cursor, for $60 billion, with the deal expected to close in Q3 2026. The acquisition follows an April 2026 partnership that gave SpaceX the option to either pay $10 billion for model-training access or purchase the company outright.
Anthropic reported that as of May 2026, Claude authored more than 80% of code merged into its codebase, with engineers merging roughly 8 times as much code per day as in 2024. The company also noted AI is increasingly involved in experiment execution, debugging, and open-ended research assistance.
Datasette 1.0a34 adds row insert, edit, and delete functionality to the Datasette web interface, available on table and row pages. The feature was prompted by the Datasette Agent tool, which had gained SQL write support via its chat interface.
Georgi Gerganov, creator of llama.cpp, stated he has used the Qwen3.6-27B model daily for roughly six weeks for coding tasks at ggml-org, running it on an M2 Ultra or RTX 5090 with a stripped-down pi agent and a short system prompt.
Vercel launched Vercel Connect, now in Public Beta, a credential management system that replaces long-lived provider tokens stored in environment variables with short-lived, task-scoped credentials issued at runtime via OIDC identity verification.
Microsoft patched a max-critical vulnerability in its M365 Copilot platform that allowed attackers to steal 2FA codes and sensitive email data via prompt injection. Researchers demonstrated a proof-of-concept exploit using markup language and HTML tags to smuggle data in web requests to attacker-...
Gartner predicts more than 40% of agentic AI projects will be canceled by end of 2027, citing high costs and unclear business value. An MIT study found enterprise AI failure rates reaching 95%, with analysts attributing the gap to AI agents lacking access to unstructured data, which accounts for ...