Welcome back. Open-source is having a massive week. From new coding models with wild multi-agent capabilities to reasoning models that match frontier performance, devs and builders are truly spoilt for choice.

On top of that, a self-hosted AI assistant called Clawdbot just hit 60K GitHub stars — proof that developers want AI they can actually control.

Also: How to use Clawdbot, how to give tough feedback without being hated, and how to start using AI for development.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Moonshot AI claims open-source coding crown: Kimi K2.5 is making waves with SOTA coding and vision capabilities, plus a novel "agent swarm" feature that coordinates up to 100 sub-agents in parallel. The model excels at front-end development and can generate code directly from images or video. Moonshot also released Kimi Code, a coding agent that integrates with VS Code, Cursor, and JetBrains. Its standout feature: visual self-debugging that spots UI errors and fixes them automatically.

Clawdbot (now Moltbot) emerges as open-source AI assistant for local control: Austrian developer Peter Steinberger — who sold his last company for €100 million — launched what he calls "Claude with hands." Originally named Clawdbot, the self-hosted AI agent was quickly rebranded to Moltbot after Anthropic raised trademark concerns over "Clawd." It runs locally on your machine and can manage files, control browsers, and message through WhatsApp and Slack. With nearly 60K GitHub stars, it's proof that developers want AI that actually does work.

Alibaba's Qwen drops new reasoning model and tough agent benchmark: The Chinese e-commerce giant just launched Qwen3-Max-Thinking, a reasoning model matching GPT-5.2-Thinking and Claude-Opus-4.5 across 19 benchmarks. Notable features include: adaptive tool-use that autonomously triggers Search, Memory, and Code Interpreter. Users also got DeepPlanning, a benchmark testing long-horizon agent planning with real-world constraints like travel scheduling and coupon optimization.

Most AI notetakers are just awkward meeting bots, transcribing what was said and sending you a summary.

Granola is an AI notepad, and there’s a huge difference.

Perfect for back-to-back meetings. 

Try Granola for a month at no cost with code: THECODE (1 month off any paid plan)

Why senior engineers let bad projects fail: Senior Staff Engineer Lalit Maganti shared hard-won lessons from his time at Google: sometimes watching a bad project fail is the strategic move. The key is treating influence like currency — nitpicking code reviews costs $5, but challenging a VP's initiative could drain $50,000. Save your credibility for battles where you can actually change the outcome.

Anthropic CEO warns AI could reshape engineering teams within years: In a new essay, Dario Amodei reveals AI is already writing "almost all" of the code for some Anthropic engineers — and predicts 50% of entry-level white collar jobs could be disrupted in 1–5 years. His advice for engineering leaders: focus on "innovation" over "cost savings" when deploying AI, and start planning now for how to reassign and support employees through the transition.

How to give tough feedback without being hated: Former VP of Engineering Lara Hogan just dropped a guide for managers dealing with stuck employees. Her advice: don't just tell people to stop a behavior — give them something to start doing instead. Lead by acknowledging their concerns, stay focused on the future, and keep your message to one or two sentences max.

How to use Clawdbot: Devs can learn to self-host Clawdbot, an autonomous AI agent, on a virtual server. The guide covers terminal-based installation, API configuration, and WhatsApp integration. You will learn to build a personal assistant capable of automating real-world tasks like scheduling and reservations.

How to nail behavioral interviews as software engineers: After conducting 1,000+ interviews at Meta, Austen McDonald says behavioral rounds are where your level gets determined — not coding tests. He recommends building a "story catalog" of 3-4 high-impact projects, using the CARL framework (Context, Actions, Results, Learnings), and always leading with your largest-scope examples.

How to start using AI for development: This tutorial talks about how devs can master AI coding by practicing on known problems rather than just difficult ones. Key lessons include limiting context via guide files instead of dumping full codebases, fixing environment errors to prevent AI confusion, avoiding complex configurations, and using "Plan Mode" for reliable results.

Anthropic reveals AI safety risk from Benign Outputs: In their new study, Anthropic prompted safeguarded frontier models like Claude 3.5 Sonnet with innocent chemistry questions, then used those responses to fine-tune open-source models such as Llama 3.3 70B. The fine-tuned versions closed about 40% of the performance gap to unrestricted top models on tasks like mustard gas synthesis, far better than training on textbooks.

Rethinking the value of multi-agent workflow: Many advanced AI systems use "teams" of agents to solve complex tasks, but typically every agent is actually the same model given different instructions. This research proves that a single agent can effectively "role-play" an entire team by itself, matching the performance of complex groups while drastically cutting costs.

NVIDIA and Stanford show models gain accuracy by treating each test problem as a learning environment: Most AI stops learning after training, which limits performance on complex, unique tasks. This research lets models continue training during the test, updating themselves in real-time to discover better solutions for hard problems in math and science.

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