OpenAI to let devs publish apps to ChatGPT

Welcome back. OpenAI is quietly overhauling its developer ecosystem with "Hermes," a new initiative designed to let you ship agent workflows directly to ChatGPT's massive user base. If you're a builder, this is a massive opportunity to ship your product or feature to 800 million+ users.

Also: How to build an AI agent with Gemini 3, why development slows down, and OpenAI's forward deployed engineering playbook.

Today's Insights

New features and model for devs
5 coding skills that matter in the AI age
How to use Claude Skills for coding
Trending social posts, top repos, new research & more

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.

THIS WEEK IN PROGRAMMING

Agent Builder is getting a 'Publish to ChatGPT' option. Source: TestingCatalog

OpenAI preps new ways for devs to publish apps to ChatGPT: The AI giant is building tools that will let developers publish agent workflows directly to its flagship product. A new feature codenamed Hermes will add a button in the Agent Builder interface to push workflows straight to users. The company is also working on an Apps dashboard and a new apps directory (think GPT Store 2.0) where developers can create, manage, and distribute their tools.

Google's Antigravity code editor vulnerable to data theft: Security researchers have demonstrated how Google's new agentic code editor can be manipulated to steal credentials and sensitive code. Google has acknowledged the risks with an onboarding disclaimer.

Startup claims it trained frontier model without Big Tech resources: Silicon Valley-based AI startup Prime Intellect just unveiled INTELLECT‑3, a 106B parameter 'mixture of experts' model that outperforms many larger models on math, code, and reasoning benchmarks, trained on just 512 H200 GPUs over two months. The recipe is fully open‑sourced.

TRENDS & INSIGHTS

What Engineering Leaders Need to Know This Week

Why effective engineering leaders never stop coding

Why effective engineering leaders never stop coding: Uber Distinguished Engineer Joachim reveals the secret to surviving promotion politics: "If you're not writing code, you're not a software engineer." The key isn't delegation, but "doing stuff" by solving the unglamorous, messy problems yourself to build the trust needed to scale your influence.

Why development slows down: Software projects start fast but slow down as they accumulate complexity and technical debt. Each new feature reduces future options by making the codebase harder to change. Kent argues the solution is to alternate between adding features and restoring flexibility through code cleanup.

Why do senior engineers struggle to build AI agents: Traditional software engineering removes ambiguity, but that mindset backfires when building AI agents. Philipp Schmid, a senior Staff engineer at Google DeepMind, argues that veterans often struggle more than juniors because they try to "code away" the probabilistic nature of LLMs.

IN THE KNOW

What's trending on socials and headlines

Meme of the week

Skills to Master: AI ate 80% of coding — these 5 skills are what still matter.
Industry Experts: Curious what Forward Deployed Engineers actually do? This playbook from OpenAI's Head of FDE spells it out.
AI Leverage: A high school dropout used AI to teach himself math and ML—then landed a research role at OpenAI.
Vibe Coding: Why are senior engineers squeezing more output from coding agents than juniors? This thread breaks it down.

Tencent unveiled HunyuanOCR, a lightweight end-to-end OCR model covering detection, recognition, and complex documents.
OpenAI released unified Voice mode on web and mobile with optional classic mode.
Google launched dynamic educational visuals in Gemini to help users study scientific systems interactively.
DeepSeek has open-sourced its gold medal math model.

TOP & TRENDING RESOURCES

3 Tutorials to Level Up Your Skills

Click here to watch Claude Skills tutorial

How to use Claude Skills for coding: This tutorial teaches developers how to build custom tools for Claude Code using the "MASTER" framework. It demonstrates how to transform manual SOPs into executable AI skills, allowing teams to automate complex workflows like release management while enforcing consistent quality standards.

How to build an AI agent from scratch with Gemini 3: If you can write a loop in Python, you can build an agent. This guide will walk you through the process, from a simple API call to a functioning CLI agent.

Anthropic tackles the multi-session agent problem: AI agents still struggle to maintain context across coding sessions. Anthropic's new engineering guide introduces a two-agent approach, one to scaffold the environment and another to chip away at features incrementally, inspired by how human engineers hand off work between shifts.

Top Repos

Gemini-cli-tips: This guide covers 30 pro-tips for effectively using Gemini CLI for agentic coding.
Better-agents: It supercharges your coding assistant, making it an expert in any agent framework you choose and all their best practices.
DeepTeam: It is a simple-to-use, open-source LLM red teaming framework, for penetration testing and safeguarding LLM systems.

Trending Papers

Estimating AI productivity gains from Claude conversations: This study analyzes 100,000 real-world Claude conversations to quantify how much time AI saves on professional tasks. Results show AI reduces completion time by 80%, projecting a potential 1.8% annual boost to US labor productivity.

ToolOrchestra: This paper discusses the inefficiency of relying on massive AI models for every task. It demonstrates that intelligently routing tasks via ToolOrchestra outperforms GPT-5 while being 2.5x more efficient.

General agentic memory via deep research: This paper discusses the limitations of shallow AI memory that fails to transfer across tasks. It finds that implementing a deep research framework for encoding creates structured knowledge that significantly improves generalization and benchmark performance.