// edition · 2026-06-14
19 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
A developer describes using Claude Code and OpenAI's Codex CLI in parallel: Claude Code for interactive, exploratory coding sessions and Codex's non-interactive `codex exec` mode for scripted, repetitive tasks such as generating commit messages and version bumps.
A developer released two open-source prompt files for Claude Code: "spec," which generates a 13-section specification before coding to reduce ambiguous task guessing, and "review-audit," a read-only code audit across six axes that requires file-and-line evidence before marking any check as passed.
Anthropic's prompt caching requires explicit `cache_control: { type: "ephemeral" }` markers and charges a write premium of 1.25x (5-minute TTL) or 2x (1-hour TTL) on first requests, then 0.10x on subsequent cache hits. The 5-minute TTL breaks even at the second cache hit, yielding a 32.5% cost re...
A tutorial describes configuring "Claude Code Free," a proxy that redirects Anthropic's Claude Code terminal tool to alternative AI providers via OpenRouter, allowing use without official Anthropic authentication. The guide covers Windows installation via PowerShell and manual PATH configuration ...
A developer guide outlines seven diagnostic checks for OpenAI-compatible API base URL errors, covering issues such as mismatched API keys, incorrect `/v1` path prefixes, and wrong model names that produce 401, 404, or model-not-found errors.
A developer blog post describes a conceptual "collective semantic memory" system for AI agents on howiprompt.xyz, where vectorized records of past task failures are stored and retrieved to prevent agents from repeating the same prompt errors.
Simon Willison investigated methods to map SQLite query result columns back to their source table and column names, aiming to add this capability to Datasette. Using Claude Code, he identified three approaches: the apsw library, ctypes access to SQLite's sqlite3_column_table_name() C function, an...
Career-Ops, an open-source job search automation tool built by Santiago Fernández de Valderrama, has accumulated 53.6k GitHub stars since release. The MIT-licensed tool integrates with AI coding CLIs to scan 45+ job portals, score listings, and generate ATS-optimized resumes; its creator used it ...
A Dev.to guide describes storing OpenAI API relay configuration in three environment variables — base URL, API key, and model name — to avoid hardcoding values in source files, with Python and (implied) other SDK examples using api.wappkit.com as the relay endpoint.
GigaOm published a report arguing that production AI retrieval systems require more than vector similarity search, citing the need to combine semantic, lexical, behavioral, and business signals in a single pipeline. The report contends that tensor-native architectures handle multi-dimensional dat...
Anthropic released Claude Fable 5, which cost $9 in one developer coding test compared to $1.50 for GPT-5.5 and under $1 for GLM-5.1 for equivalent results. The US government subsequently ordered Anthropic to disable access to Fable 5 citing export-control concerns over a reported jailbreak metho...
Anthropic released Fable 5, its highest-capability generally available model and the first in its new Mythos-class tier, priced at $10 per million input tokens and $50 per million output tokens. The launch drew criticism after a disclosure in its 319-page system card revealed the model would sile...
Ollama and vLLM are two open-source tools for running large language models locally: Ollama targets ease of use for individual developers, while vLLM is designed for high-throughput production inference with support for concurrent requests at scale.
MiniMax M3 scores 59% on SWE-Bench Pro at $0.60 per million input tokens, while Claude Opus 4.8 scores 69.2% at $5.00 per million input tokens. Both models offer 1M token context windows; M3 also supports video input and has promised open weights not yet released.
A commentary in The New Stack argues that AI-assisted software development is shifting from human-driven prompts to automated feedback loops — small programs that prompt AI agents, evaluate outputs, and retry until goals are met. The author contends this shifts the developer's role toward buildin...
A developer blog post compares a claimed 1-trillion-parameter open-weight coding model called Kimi K2.7-Code, attributed to Moonshot AI, against Claude Opus 4.8 and GPT-5.5, citing a price of $0.95/$4.00 per million tokens — 5-7x lower than the competing models.
A developer published a guide outlining six strategies for early-career programmers to build skills as AI handles routine coding tasks, including contributing to real open-source projects, reviewing AI-generated code before shipping it, and building a visible track record of contributions.
Anthropic reportedly released "Claude Fable 5," a flagship model, which the U.S. Commerce Department then restricted to U.S. citizens only, citing a jailbreak; Anthropic subsequently disabled both Fable and Mythos models for all users. The article also notes Fable 5 carries a mandatory 30-day dat...
Anthropic cut off worldwide access to two of its AI models on Friday, reportedly after Amazon CEO Andy Jassy raised security concerns about them.