Welcome back. AI agents were meant to free up developers, but they might actually be doing the opposite. An AI startup founder shared a post this weekend that sparked a serious rethink of what "productivity" looks like when agents are doing the heavy lifting.

Also: How OpenClaw creator closed 4,000 GitHub issues in a day, cut Claude Code cost by 10x, and why a 110-person team woke up locked out of Claude.

Alibaba ships upgraded image model for production: The Chinese tech giant just dropped Qwen-Image-2.0-Pro, boosting image quality and text rendering while keeping visual styles consistent. It’s already ranked #9 on the Arena text-to-image leaderboard, holding its own against OpenAI and Google, and #17 for single-image editing. Devs can access it through ModelScope or the Alibaba Cloud API. Try it here.

Google pours $40B into Anthropic to scale Claude: The search giant just committed $10 billion upfront, with another $30 billion tied to performance milestones, valuing Anthropic at $380 billion. This investment follows a massive 5-gigawatt compute deal with Broadcom earlier this month. The funding helps scale Claude as demand grows, while Google Cloud continues to offer Claude on its TPUs as an alternative to Nvidia hardware.

OpenAI publishes a five-principle framework for AGI development: The ChatGPT creator just released its most significant update to its guiding charter since 2018, stepping back from an earlier promise to avoid competing with rivals as they near the finish line. CEO Sam Altman suggested they may prioritize system resilience over user empowerment, likely leading to stricter API guardrails as the tech evolves.

Hate to break it to you, but your LLM might be choosing patches that look clean over patches that actually work.

AI21 discovered this error by testing their own SWE-bench agent. Even on fresh, post-cutoff data (SWE-rebench), LLM judges consistently preferred minimal patches, even when they were wrong.

But it’s not contamination, it’s something they call "gold-like bias”-

and it could be affecting your evals, too.

Read AI21’s study to detect, isolate, and eliminate this bias gap completely.

The signal is fading. Technical interviews have fundamentally shifted this year. Candidates are using cheating tools that feed them real-time, AI-generated answers directly on their screens. Since these tools are invisible to screen-sharing software, hiring managers are struggling to catch them. Consequently, the technical round has lost its status as the ultimate filter.

The round that decides offers. This shifts the spotlight to the behavioral round, which is now what actually lands the job. Steve Huynh, a former Amazon Principal Engineer with nearly a thousand interview loops under his belt, notes that candidates rarely fail on their code. Instead, they get rejected because they can’t tell a clear or compelling story about their own experience.

A widening gap. Most engineers obsess over coding prep but wing the behavioral round at the last minute. Now that AI has leveled the playing field for technical skills, your ability to articulate your thought process is your only real competitive advantage.

Put in the work. Spend ten hours recording yourself answering basic questions like "tell me about yourself" or "why do you want to work here?" Watch the recordings to trim the fat and do the same for three stories about tough career calls. For a full breakdown on how interviewers score these sessions, check out the deep dive by The Pragmatic Engineer.

NullPointerException in checkout-service. Viktor read the Sentry stack trace, pulled Datadog logs, diffed the last six commits on GitHub, and found the bug: a recent refactor dropped a null check on the customer object. Viktor opened a PR with a one-line fix and a test. Filed a Linear ticket. Tagged the engineer who shipped it. Posted to #on-call: "Isolated. Customer impact: 3%. PR needs review before deploy."

Sentry, Datadog, GitHub, Linear, PagerDuty, and 3,000+ more. Your data never trains models.

Start free. $100 in credits →

Offline Stack: A bootstrapped founder who exited at $50M ARR replicated agentic coding entirely on his laptop. He shared the full model, agent, and sandbox setup.
_{1,200 bookmarks}

Issue Sweeper: OpenClaw's creator, Peter Steinberger, built ClawSweeper, running 50 codex agents in parallel, and just closed 4,000 GitHub issues in one day.
_{2 million views}

Ban Radius: One employee's behavior got an entire 110-person company locked out of Claude overnight. Every founder running Claude through their org should read this.
_{2,800 likes}

Fork Flag: This guide shows a hidden Claude Code environment variable that makes subagents 10x cheaper while keeping your context clean.
_{1,300 bookmarks}

Custom Evals: A Google AI Studio engineer argues every AI company should be building their own benchmarks, or risk model progress passing them by.
_{1,100 likes}

Long-running Claude Code tasks can be a total productivity killer. You kick off a process, switch tabs, and inevitably end up checking back way too late or wasting time hovering over the terminal. Adding sound cues is an easy fix that keeps things interesting.

To set it up, just drop your “.wav” files into “~/.claude/hooks/” and run this command in Claude Code:

Claude automatically updates your “settings.json” and configures the hooks for you. You can swap out audio files whenever you want to change your sound pack.

If you prefer a manual setup (like Anthropic engineer Delba did with her Warcraft peon sounds), you can find her full configuration here.

How to use Codex to accelerate real-world coding workflows: This FreeCodeCamp tutorial teaches developers how to leverage Codex for coding and productivity. You'll master the CLI, session management, and context window control, while exploring advanced features like plan mode, custom skills, sub-agents, and sandboxes to automate your entire workflow.

Factory: An AI-native software development platform that works wherever you do. Delegate entire tasks, like refactors, incident response, and migrations, to Droids (agents) without changing your tools, models, or workflow.

Agent skills for real engineers (26.4k ⭐): Ex-Vercel engineer Matt Pocock open-sourced the personal Claude skills directory he relies on daily for actual engineering, rather than just "vibe coding."

Autogenesis: Current AI agents have a hard time adapting safely because they don't have a standardized way to manage their own updates. Autogenesis fixes this by treating agent components as trackable resources, allowing for reliable self-improvement and better performance on complex tasks.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 250K+ engineers and 150K+ 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