Self-improving Claude agents via verification loops and CLAUDE.md
Anthropic engineers reportedly build self-improving agents by closing the loop with self-verification and persistent CLAUDE.md context.
How DeepSeek Runs a 284B LLM on a Laptop (Run AI Locally)
DeepSeek's 284B MoE model runs locally at ~26 tokens/sec on a 128 GB MacBook via a pure-C engine and asymmetric quantization: 2-bit experts, 8-bit router.
Anthropic adds ID verification to Claude via Persona
Anthropic rolls out identity verification on Claude using Persona. ID and selfie are held by Persona, not Anthropic, and not used for model training.
Claude Tag: Slack-native team threads with ambient mode
Anthropic released Claude Tag, a Slack integration that lets whole teams collaborate with Claude in shared threads, with optional ambient monitoring.
Codex Record and Replay turns a demo into an automation
Codex's new Record and Replay watches a workflow once and turns the demonstration into a repeatable automation, collapsing demo-to-SOP into a single motion.
Oak: open-source VCS designed for AI coding agents
Oak is a new open-source VCS in Rust designed for AI coding agents, replacing git's per-commit model with branch-per-session and content-addressed lazy mounts.
Why Smarter AI Routing Is the Cost Advantage Nobody Is Talking About
Ramp CEO Eric Glyman frames the AI cost curve as bifurcated: basic tasks are 700x cheaper, frontier tasks 300x more expensive. Smart token routing captures t...
Claude Code's underused /goal command for autonomous agent loops
Claude Code's /goal command runs an autonomous loop: take an action, check the goal condition, repeat until met, replacing step-by-step prompting.
LLM Evals: Everything You Need to Know
A sharp-opinion FAQ on LLM evals: skip generic metrics and tooling, do error analysis on your own traces, build a failure taxonomy, then write targeted evalu...
Reflexion: Language Agents with Verbal Reinforcement Learning
Reflexion lets an LLM agent learn from failures without fine-tuning, writing a verbal post-mortem and feeding it back as context, lifting GPT-4 to 91% on Hum...
Supacode: terminal-first agent harness manager for macOS
Supacode is a terminal-first agent harness session manager on libGhostty for macOS, built to manage many AI coding agents and worktrees at once.
Claude Code Live Artifacts: auto session summary to share
A short clip claims Claude Code now auto-generates a Live Artifact summarizing code and PRs at the end of a session to share with teammates.
Claudoro: Pomodoro timer embedded in Claude Code
Open-source Pomodoro timer that lives in the Claude Code status line, with an immutable JSONL log and zero-token CLI so the timer never costs API tokens.
Pagecast: publish HTML or Markdown from your coding agent
CLI tool that lets a coding agent like Claude Code or Codex publish HTML or Markdown to your own free Cloudflare Pages account with one confirmation.
A Practical Guide to SSH Tunnels: Local and Remote Port Forwarding
Mnemonics and labs for the four SSH tunneling modes: local (-L), remote (-R), dynamic (-D), and remote-dynamic. Covers bastion hosts and SOCKS proxies.
Eloquent: Google's free on-device AI voice dictation app
Eloquent is Google's free voice dictation app that turns rambling speech into clean text, running on-device so it stays private and works offline.
Loop Engineering: Run Coding Agents in Iterative Loops
A short teaser for loop engineering, the idea that AI coding agents work best run in an iterative loop rather than expecting one-shot results.
Mega-ASR: open-source ASR built for noisy real-world audio
Open-source (Apache-2.0) foundation speech-recognition model built for messy audio, claiming up to ~30% WER gains over SOTA where other models collapse.
Local Qwen isn't a worse Opus, it's a different tool
Local Qwen 27B isn't near-Opus for coding, but it pays off on privacy-bound analysis where cloud models are contractually off-limits.
ralphex: The Extended Ralph Loop for Claude Code and codex
Open-source MIT orchestrator that drives coding agents through a structured plan, running each task in a fresh session plus multi-agent code review.