Welcome back. Software engineering best practices were built for humans. Now that agents write most of the code, we're rebuilding those practices from scratch. Anthropic's new Monitor tool for Claude Code is a perfect example. See how it works.

Also: How to build an agent harness, 50+ guides to ship with Claude, and what Karpathy thinks separates AI hype from reality.

Claude Code gives developers a security assistant: The AI lab just dropped Monitor tool, an update that lets its coding agent run background scripts to track real-time events. It is basically a security camera for your code that only alerts you when it spots movement. You can set Claude to watch for dev server errors, track failing tests, or monitor a production launch for hours. This approach saves tokens because it only fires up when something actually goes wrong.

Cursor brings screenshots and video to agent-generated PRs: The AI coding startup just shipped a feature that automatically attaches demo videos and screenshots to the GitHub PRs created by its cloud agents. This lets teams visually review AI-generated code changes without ever leaving the platform. In one demo, an agent built a color picker, refined the UI, added tests, and dropped a 20-second clip of the finished feature directly into the PR.

OpenAI adds $100 Pro plan for heavier Codex use: The ChatGPT maker just rolled out a new $100/month Pro tier for developers who have outgrown the standard $20 Plus plan. It offers five times the Codex usage, an exclusive Pro model, and unlimited Instant and Thinking models. Designed for heavy coding sessions, the plan includes a launch promo where new subscribers can access up to ten times the standard Plus-level usage through May 31st.

For 10+ years, AI Council has been gathering the world's top AI infrastructure minds to share what’s working (and where we’re headed next). 

You’re invited for 3 days of high-quality discourse with 1,200+ technical experts, including office hours, small groups, and zero marketing keynotes. 

Speakers include:

…and more. Takes place May 12-14 in San Francisco.

See the full speaker lineup here and use code THECODE20 for 20% off.

The reasoning era is over. Junyang Lin, the former tech lead behind Alibaba's Qwen models, stepped down in early March after what colleagues called an involuntary departure. A few weeks after leaving, he published his first essay with the blunt thesis that the reasoning wave sparked by OpenAI’s o1 and DeepSeek’s R1 is already winding down, only to be replaced by agentic thinking.

Interaction over monologue. While reasoning models just think longer before they speak, agentic models think in order to act. They use tools, read environment feedback, and pivot mid-task. The key question is no longer "how long can it think?" but "can it execute effectively over multiple turns?" This shift requires models that know when to stop ranting about theory and start doing.

More thinking doesn't mean more intelligence. Recent research shows that simply increasing test-time compute doesn't always yield better results. In fact, giving agents tool access makes training riskier, as they can exploit logs or search for answers rather than actually solving problems.

Environments are the new moat. Lin points out that building better environments is becoming its own startup category because training agents now involves browsers, terminals, and API layers instead of just models. Your competitive edge is shifting from RL algorithms to harness engineering and the ability to close the loop between what a model decides and what actually happens next.

Claude Code loads your CLAUDE.md at session start. But it doesn't know what branch you're on, what files changed, or what broke since your last session. You end up typing "check git status" before doing anything useful.

A hook pattern from Claude Code's hooks system fixes this for good. Just add this to your “.claude/settings.json”:

That's it. Now, whenever you start a session or clear your history, Claude automatically pulls in your current branch and any changes. You'll have all the context you need to get straight to work.

How a senior software engineer uses coding agents: This tutorial shows you how to turn Claude Code into a powerful autonomous assistant. You'll learn how to set up project context files, build automated validation loops to catch bugs, run parallel sub-agents, and master a strict "plan, execute, and review" workflow to supercharge your coding.

Archon: A harness builder for AI coding agents. You can define your entire development process from planning and implementation to code reviews and PRs, as YAML workflows that run reliably across all your projects.

ConvApparel (by Google): Today's AI user simulators feel way too perfect and unnaturally patient compared to real people. But when you train these simulators on actual human data instead of just feeding them instructions, they become much more realistic and adaptable.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team