Welcome back. Trying to build an app over the last decade kinda sucked if we’re being honest. SEO was saturated, Facebook and Google ads became too expensive, and unless you raised a ton of venture capital, finding distribution for your products was really, really hard. That’s changing now.

Today: OpenAI claims GPT-5 is less politically biased, how to use AI for software engineering tasks, and OpenAI’s tutorial on multi-agent coding workflows

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.

Google and AWS race to unify enterprise AI: Google Cloud launched Gemini Enterprise to consolidate workplace AI tools with no-code agents and cross-platform data access. Meanwhile, AWS debuted Quick Suite, connecting 1,000+ apps for automated workflows, research, and data visualization across organizations.

New distribution channels emerge for app developers: First, OpenAI's new Apps SDK turns ChatGPT into a distribution platform where over 800M users can discover apps mid-conversation. Second, as traditional search loses ground to conversational AI, smart developers are optimizing for citations by LLMs using GEO (Generative Engine Optimization). Here's a brief guide for developers on GEO implementation. Third, X announced it's testing timeline changes to give external links equal treatment, reversing years of deprioritization that hurt app makers and publishers.

OpenAI reduces GPT-5 political bias by 30%: The ChatGPT maker just released new research showing that its GPT-5 models exhibit 30% lower political bias than previous models. Using 500 test prompts across 100 topics, researchers found that GPT-5 stays remarkably neutral on balanced questions but can show moderate bias when hit with emotionally charged prompts.

How to use AI to help with software engineering tasks: A new prompting framework promises to help software teams unlock AI's full potential for everyday development work. The framework emphasizes placing context first to "prime" the AI model, then layering in role definition, specific actions, output formatting, communication tone, concrete examples, and final constraints.

Why most AI agent projects fail before a single line of code is written: A new playbook by agentic AI practitioners Sara Davison and Tyler Fisk reveals that teams often build agents around documented processes rather than the realities of how work actually gets done. They argue that successful agents must be grounded in an understanding of four hidden layers of work intelligence: surface procedures, operational reality, contextual intelligence, and cultural DNA.

Firefox CTO shares hard-won lessons from the browser wars: Bobby Holley climbed from intern to CTO of Mozilla Firefox over 17 years, living through the brutal competition with Chrome. His biggest career insight: focus relentlessly on impact rather than chasing promotions.

Cursor Agent —10 Pro Tips: A new tutorial reveals how developers can supercharge their coding workflow with Cursor's AI agent. It covers tips to use Plan Mode, creating custom slash commands for repetitive tasks (like automated PR generation), passing images directly to the agent for visual design matching, and using branch-level code reviews.

Building multi-agent coding workflows (Tutorial): OpenAI published a detailed tutorial demonstrating how to orchestrate teams of AI agents for software development. They built a five-agent system where specialized AI developers collaborate like a real software team — with a Project Manager coordinating handoffs between Designer, Frontend Developer, Backend Developer, and Tester agents.

Agents 2.0: From Shallow Loops to Deep Agents: A senior AI engineer at Google DeepMind has shared a new architectural pattern for building AI agents. The architecture, called “Deep Agents,” separates planning from execution through four core components: explicit planning tools, sub-agent delegation, persistent memory, and sophisticated prompt engineering.

LLMs spread more lies when competing for audiences: Stanford shows LLMs competing for audiences can boost sales or engagement, but simultaneously increase deception, disinformation, and harmful rhetoric, revealing competitive pressures systematically erode alignment safeguards.

LLMs Reproduce Human Purchase Intent: This paper shows that you can predict actual purchase intent (90% accuracy) by asking an LLM to impersonate a customer with a demographic profile, giving it a product & having it give its impressions, which another AI rates.

Agent Learning via Early Experience: Meta researchers propose a new learning paradigm for language agents called "early experience"—a reward-free method where agents learn by interacting with environments using their own suboptimal actions, instead of relying solely on human demonstrations or reinforcement signals.

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

Until next time — The Code team