Welcome back. Coding agents are smart, but they've got the memory of a goldfish. Every time you start a new session, you're stuck re-explaining your entire project from the ground up. Anthropic's latest update to Claude Code just fixed that once and for all.

Also: Using Claude Code to scrape websites, a developer's playbook to build bootstrapped businesses, and why Block just laid off 4,000 people.

Claude Code gets persistent memory across sessions: Anthropic just rolled out auto-memory for Claude Code, allowing the AI coding agent to save project context, debugging patterns, and workflow preferences across sessions. CLAUDE.md now serves as instructions for Claude, while MEMORY.md acts as a scratchpad that Claude updates itself, including any specific details developers ask it to remember. See how it works.

Google unveils Nano Banana 2, its fastest image generation model yet: The search giant just shipped Nano Banana 2, a new image model that combines the advanced capabilities of Nano Banana Pro with the speed of Gemini Flash. Key updates include consistent character rendering for up to five subjects, 4K resolution support, sharp text generation, and much better instruction following. Watch Nano Banana 2 in action.

Anthropic refuses Pentagon's demand for unrestricted AI access: CEO Dario Amodei turned down the Pentagon's demand for unrestricted access to Claude, standing his ground on two key issues: no mass surveillance of Americans and no lethal autonomous weapons. While OpenAI and xAI have reportedly already agreed to the new terms, this standoff could create problems for engineering teams that rely on Claude for defense and government projects if the Pentagon decides to cut ties with Anthropic.

You could have the most intelligent AI in the world, but if it’s slow, your users will leave.

Whether you’re already in the dev phase or just brainstorming, the Latency in Conversational AI eBook can help you diagnose your latency issues before it’s too late.

Learn how top teams balance speed and quality by…

Don’t settle for “good enough.” 

Download the eBook and make your AI exceptional.

The old math is broken. For decades, engineering rituals like sprint planning and scoping existed because writing code was expensive. Coding agents have slashed those costs to nearly zero, yet most engineering teams haven't updated their processes to match. Simon Willison, co-creator of Django, recently shared a guide on agentic engineering patterns that outlines exactly what needs to change.

Override your instincts. Willison’s main point is simple: whenever you feel like something "isn't worth the time to build," just run a prompt anyway. Kick off an asynchronous agent session — at worst, you'll spend ten minutes checking the result before discarding it. Agents can now handle refactoring, documentation, and test coverage in minutes, tackling tasks that used to take all afternoon.

Let the agent prove its own work. One of the most effective patterns Willison points out is red/green test-driven development. You have the agent write a failing test first, and then you have it write the code to pass that test. This forces the agent to prove its work before you even bother looking at it, which helps you avoid that common headache where code looks solid but doesn't actually run.

The clock is already running. The shift isn't just changing how engineers work — it's changing how teams hire. One startup recently redesigned its entire interview loop around AI-assisted coding, choosing to test for architectural judgment over raw implementation speed. Teams that haven't rewired their process will notice the gap soon enough.

You build the app, launch it, and land your first enterprise customer. 

Then security steps in: SSO, roles, audit logs, abuse protection, etc. You’re not ready. Growth stalls while you duct-tape infrastructure. You lose the client.

Or, do what top AI players do. WorkOS gives you the invisible stack you need—enterprise auth, RBAC, multi-tenancy, provisioning, audit logs, real-time protection—without rebuilding from scratch.

It’s the plug-and-play solution that passes all the tests, so you become enterprise-ready in days (not years). 

See WorkOS in action here.

Claude Code's plan mode has a gap. You approve a plan, Claude starts building, and before you know it, you’re looking at hundreds of lines of code filled with missing tests, scope creep, and over-complicated abstractions.

Y Combinator CEO Garry Tan shared a slash command that fixes this. His “/plan-exit-review” command forces Claude to plan and critique its approach before writing anything.

To set it up, copy the full prompt from his GitHub gist and save it as:

Then, after using plan mode, run:

Claude reviews its own plan across four key areas, flagging broken file references and providing numbered fixes for each. It also runs a scope check right out of the gate to kill any busywork before it hits your codebase.

How to set up your codebase for agents: This tutorial shows you how to prep your code for agentic coding. The secret is using "deep modules" that hide complex logic behind simple interfaces. By applying classic software design principles, you make your codebase much easier for AI to navigate and update.

Scrapling: A smart web scraping framework that handles everything from simple requests to large scale crawls. It automatically adapts to website changes, bypasses common anti-bot systems, and lets you run fast, scalable crawls with just a few lines of Python.

Do current agentic memory systems scale? AI agents need memory to handle long term tasks, but current testing methods are flawed and tend to overlook hidden performance costs. This paper suggests that the real hurdle isn't just coming up with new architectures, but actually creating solid evaluations and making sure these systems run smoothly and efficiently without breaking down.

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