OpenAI's Experiment: Shipping a Real Product with 100% Agent-Written Code
Today's top 24 insights for PM Builders, ranked by relevance from Blogs, X, YouTube, and LinkedIn.
OpenAI's Experiment: Shipping a Real Product with 100% Agent-Written Code
#1 š OpenAI News
Harness engineering: leveraging Codex in an agent-first world - An engineering post describing how the Codex harness is used to enable an agent-first world, explaining engineering approaches and system design decisions. It highlights practical implementations and lessons learned while integrating Codex into agent workflows.
#2 š
Google DeepMind shows how Gemini Deep Think uses agentic workflows to accelerate solving research-level problems in mathematics, physics, and computer science in two new joint papers with Google Research.
#3 š
NVIDIA AI introduces Blueprint for Video Search and Summarization (VSS) and the open VLM Cosmos Reason in its GTC session. These new features empower AI agents to reason over video and extract actionable insights across edge, on-prem, and cloud.
#4 š
Andrej Karpathy stripped nanoGPTās entire architecture and loss down to atomic operations (+, *, **, log, exp). He then uses a tiny scalar autograd engine (micrograd) to compute gradients and Adam for optimization.
#5 ā¶ļø
Claude Opus 4.6 vs GPT-5.3 Codex: How I shipped 93,000 lines of code in 5 days
How I AI Podcast
Head-to-head testing of OpenAI GPT-5.3 Codex in Codeex and Anthropic Opus 4.6 (plus Opus 4.6 Fast) in Cursor to redesign a PLG+enterprise marketing site and refactor core application components, resulting in 93,000 lines of code shipped in five days.
- In five days, the team merged 44 pull requests containing 98 commits across 1,088 files, adding 92,000 lines of code and removing 87,000 lines for a net addition of 5,000 lines.
- Opus 4.6 Fast is priced at about $150 per million output tokens, six times the cost of standard Opus 4.6.
- GPT-5.3 Codex in the Codeex desktop app introduces Git primitives (branches, work trees), built-in skills, and scheduled automations as first-class features.
#6 ā¶ļø
Claude Code Built My $450K Marketing Campaign
Greg Isenberg
Jonathan Courtney used WhisperFlow and Claude to transcribe a 15-minute audio brain dump into an HTML āPromoter Blueprint,ā then migrated it into Claude Code to build and deploy a Vercel-hosted marketing web app targeting 4,000 webinar signups and $450,000 in revenue.
- Imported a 15-minute brain dump via WhisperFlow into Claude and generated an ADHD-friendly scannable HTML document of podcast talking points.
- Converted the HTML into a .claudemd context file in one hour, then created a Claude Code project folder and deployed the interactive āPromoter Blueprintā on Vercel via GitHub.
- Compiled a 17,766-line document of marketing emails into Claude for best-practice extraction, feeding an email/webinar campaign designed to convert 4,000 signups into $450,000 revenue.
#7 š
Santiago built On-Call Health, an open-source tool that merges objective signals (PagerDuty, Slack, GitHub, Linear, Jira) with quick self-report surveys to compute a burnout risk score and flag engineers needing proactive check-ins.
#8 š
Boris Cherny launched an opt-in open-source sandbox runtime for Claude Code that runs locally with file and network isolation to boost safety and cut permission promptsājust run /sandbox to enable (Windows support coming soon).
#9 š
Guillermo Rauch announced Vercel Flags, a built-in feature-flag system on Vercel that helps teams ship fast without breaking things. Itās designed to de-risk agent-driven engineering by controlling feature rollouts as your team scales.
#10 š
DeepLearning.AI launched a new short course on the A2A Agent2Agent Protocol, standardizing cross-framework agent discovery and communication without custom glue code.
#11 ā¶ļø
Learn The AI Agent Cron Job Inception Strategy (Claude Code)
All About AI
The video demonstrates an AI Agent Cron Job Inception Strategy using Claude Code on a Mac Mini, where cron jobs in chron jobs.json can spawn one-time chrome jobs via rules defined in a claw MD file to autonomously discover and schedule new tasks.
- Explore-and-spawn rules in the claw MD file allow any chrome job to spawn a one-time chrome job with constraints: self-deletion after execution, a minimum 15-minute delay before execution, a maximum of two spawn jobs per session, and a mandatory spawn evaluation at the end of each job.
- Claude Code on the Mac Mini initially loads 24 scheduled tasks from chron jobs.json and logs execution via a background script, then uses spawn evaluations to append new cron jobs autonomously.
- In a test run, the agent read the Hacker News front page, spawned a chrome job scheduled at 15:57 to analyze the GitHub repo ātotal recall gated memory for cloud codeā and post a comment on its Hacker News discussion, then chain-spawned a follow-up chrome job at 16:25 to reply on X to user āDave Goldbot.ā
#12 š
Aravind Srinivas integrated Memory with Model Council, letting you run multiple frontier reasoning models on your data in concert for richer insights (video demo included).
#13 š
Santiago shows how Claude Codeās new /insights command auto-generates a detailed report. Attach that file to a Claude session and get a 10-item bullet list of lessons to supercharge your Claude Code workflow.
#14 š
Qwen patched Qwen-Image 2.0 in Qwen Chat to fix classical Chinese poem ordering and character consistency in image generation and editing. The update is live now for testing.
#15 š
Claude now includes file creation, connectors, and skills in its free tier, unlocking these core features for all users without a subscription.
#16 š
LlamaIndex š¦ LLM-powered coding agents are disrupting open source by flooding projects with low-quality AI-generated PRs, stunting developer skill growth, and turning knowledge sharing into an AI-fronted experience.
#17 š
NVIDIA AIās Nemotron Parse table extraction model is now integrated into DocuSign, enabling more accurate, automated document workflows.
#18 in
Tal Raviv argues that YCās top askāāCursor for PMsāāis already within reach if Cursor strips git from its UI, adopts a Notion/Obsidianāstyle editor, and surfaces MCP integrations (ideally as a separate ānormie friendlyā mode).
#19 in
Dharmesh Shah reports HubSpotās Q4 2025 added 9,800 new companies, pushing the total to over 280,000 on its agentic customer platform. He argues the best companies devote every calorie to delivering real customer value, not āvibe codingā around core platforms.
#20 in
Udi Menkes: Levelsioās new project OpenClawāa landing page selling Cloudbot skills sourced from X and Redditāhas already earned $12.93 and spurred teams to deploy secure, organizationāwide GenAI agents performing work at a 20āemployee scale.
#21 š
Claude added Compaction to its free plan, automatically summarizing earlier context so you can maintain seamless long conversations without restarting.
#22 š
Andrej Karpathy released a 243-line, dependency-free Python implementation of GPT that handles both training and inference. He asserts this script captures the full algorithmic coreāeverything else is just performance tuning.
#23 ā¶ļø
Use A2A to connect agents across different frameworks and teams
Deeplearning.ai
The course teaches how to use A2A: the open Agent2Agent Protocol to standardize discovery and communication between AI agents built with different frameworks via a client-server model.
- A2A was announced by Google Cloud in April 2025 and donated to the Linux Foundation in June 2025
- IBMās agent communication protocol ACP merged into A2A shortly after its donation
- The course builds three remote A2A agents: an insurance assistant with no framework, an LLM research agent using Google Agent Development Kit, and a doctor matching agent with LangGraph
#24 š
Boris Cherny says Claude Code works great out of the box but offers deep customizationā37 settings and 84 env vars via settings.json at codebase, folder, user, or enterprise scope. He recommends checking your settings.json into git so your whole team can benefit.