Deepagents separates agent loops from filesystem-like backends

Riley Brown uses Codex skills to research YouTube videos, generate Remotion graphics and Excalidraw diagrams, draft Notion content outlines, and build Paper-based YouTube thumbnail variations for his 1.5M+ follower AI-content business.

• His YouTube researcher skill uses the Supadata API to pull a full transcript from a YouTube video in about one second; Codex sub-agents can scrape an entire channel in about 30 seconds.
• He chains an “internet image puller” skill using SerpAPI and Google Images with a Remotion best-practices skill: the agent pulls relevant logos from a video transcript, creates branded graphics, and exports them as overlays for edited video.
• For thumbnails, Codex scrapes high-performing thumbnails—including Alex Hormozi and Dan Martell examples—into Paper, then Riley Brown uses Paper’s built-in AI image generation to replace the original subject with his own image and iterates on details such as face smoothing, outline glow, text, and colors.

Ian Silber describes how OpenAI’s product-design team uses ChatGPT, Codex, and ChatGPT Work to prototype ideas, test durable interface decisions, and ship experimental product changes quickly.

• Ian Silber said engineers have sometimes increased productivity by 10x to 100x with AI, while design work still requires repeated feedback loops, including trying ideas, discarding them, and gathering user or internal feedback.
• For some ChatGPT features, OpenAI tries about 100 options, discards 99, and ships one; for other work, teams use a “building in public” approach, take large swings, and learn from rapid feedback.
• Ian Silber uses ChatGPT Work, which runs in the cloud, and Codex to turn early ideas into prototypes or visual artifacts, including from a phone when he is away from a computer; he also uses AI to summarize Slack follow-ups and prepare context for meetings and recruiting.

A machine-learning model trained on 5,335 resolved Kalshi Bitcoin up/down 15-minute markets used the first 5 minutes of each market to place directional trades with 10 minutes remaining, producing 13 wins in 14 $5-stake entries but an estimated approximately -$12 net P&L after fees over 1,000 windows.

• The dataset contained 5,335 resolved markets, with 600 markets held out untouched for final testing; the model was trained using the first 5 minutes of each 15-minute window, from T900 market opening to T0 resolution.
• The bot retrieved Kalshi data and candles, selected Yes or No at the 5-minute decision point, and initially placed $1 orders; its first 16 live trades recorded 12 wins and 4 losses, a 75% directional win rate.
• After increasing the stake to about $5 per entry, the recorded performance was 13 wins out of 14 entries, 93% win rate, +$31 net P&L after fees, and 45% return on money spent; the balance rose from $56 to $88 during the run.