CLAUDE CODE

The andrej-karpathy-skills repo exploded because every coding agent needs behavioral rails. The useful move is not copying it blindly, but turning the rules into repo-specific operating constraints.

Efficient agents do not stuff every tool result into the model context. They keep intermediate state in code, files, and execution environments, then return compact summaries and receipts.

Manual approval prompts stop protecting users when coding agents ask too often. The better pattern is risk-aware autonomy: safe defaults, narrow deny rules, and approvals only for meaningful changes.

Claude Code is turning into an orchestration layer for agent teams. Here is how subagents, MCP, hooks, and long context fit together in 2026.

A long-running coding agent is only useful if the environment around it can queue tasks, capture logs, checkpoint state, verify behavior, limit cost, and recover from failure.

Skills turn a general coding agent into a trained teammate by packaging runbooks, scripts, examples, and domain-specific judgment into reusable instructions.

GitHub trending is full of agent skill frameworks. The real shift is not bigger prompts or more agents. It is turning team process into inspectable, reusable operating instructions.

I told an agent to improve the site every 10 minutes and went to sleep. Here is what 12 new repos, 60 PRs, and three goofs taught me about overnight orchestration.

The MCP ecosystem crossed 22,000 servers in early 2026. Most are noise. Here are the open-source servers that have earned a permanent slot in our config, with copy-paste setup for Claude Code, Cursor, and Codex.

Claude Code does not have to call Anthropic's API. Here are five working patterns for running it through your own gateway, on your own models, in your own VPC, with full audit logs and cost control.

Agent runs are opaque. TraceTrail turns a Claude Code JSONL into a public share link with a stepped timeline of messages, tool calls, and tokens.

The Claude Code skills and plugins hub: what skills are, where to find them (Anthropic's official marketplace, anthropics/skills, top GitHub repos), how the plugin marketplace works, the Superpowers methodology plugin, exact install commands, and the picks worth your context budget.

Hospitals still ship HL7 v2 pipes between systems in 2026. Here is how to wire Claude Code as a careful, HIPAA-aware migration agent that takes them to FHIR.

Claude Code hooks are powerful, but discovery and install still feel like manual JSON surgery. The Hookyard prototype shows what a hook package manager should become.

A curated directory of 312 Claude Code skills, plus Pro tools for authors who want analytics, version pinning, and a real submission flow.

A repo-root DESIGN.md gives Claude Code, Codex, and other agents the design rules they need to honor so generated UI does not drift into generic territory.

Zilliz's Claude Context MCP gives coding agents semantic code search, but the real question is whether retrieval makes agent work more reviewable, cheaper, and easier to verify.

The second half of our agent tooling release: distribution, validation, and ergonomics layered on top of the first six. Six small CLIs, one through-line.

Two quality-of-life tools we built this week for Claude Code daily drivers: a SKILL.md linter and a VS Code status bar that shows live LLM spend.

Ten private tools shipped overnight - observability, skills, hooks, prompts, and evals - aimed at the agent infrastructure gap small teams keep falling into.