DeepSeek open-sources a super-agent prototype that watches your GitHub repo and opens pull requests
•Season 3•Episode 34
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
6:22
Top News: DeepSeek open-sources a super-agent prototype that watches your GitHub repo and opens pull requests
Plus: OpenAI will put ChatGPT and Codex on one shared token allowance starting in December
Top AI Tools:
Beacon (https://beacon.sh/) gives coding agents like Claude Code, Cursor, and Codex an open-source memory layer that learns from past sessions, so they stop repeating the same mistakes.
Cedric (https://meet-cedric.com/use-cases/ai-solution-for-startups) is an AI teammate for startups that works inside Slack, Discord, and Telegram, connects to thousands of tools, and can join your Zoom, Meet, or Teams calls to present and send follow-ups.
Xmind 2027 (https://xmind.com/) turns your mind map into an AI thinking partner, with an Xmind Agent that builds on your own documents and works alongside ChatGPT and Claude.
Business AI Tip of the Day
Write Attention Rules Before You Turn On an Always-On Agent
SPEAKER_00
Welcome back to Sumo Growth AI Insights. I'm Darius, and I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead Story DeepSeek open sources a super agent prototype that watches your GitHub repo and opens pull requests. DeepSeek's harness team published a long design piece on Sunday describing a super agent that keeps context across sessions, runs several jobs in parallel, and stays online all the time. Alongside it, the lab opensourced DSH Superagent, a working research prototype under the MIT license. It's a coding agent that watches one GitHub repository, opens a work ticket when an approved teammate asks for something, handles it in its own isolated session, and replies in the original thread, for example, with a pull request. DeepSeek is clear that it's reference code, not an official product, and says much of it was written by coding agents. A web panel lists only the items that need a person, following the project's core rule. Take as little human attention as possible and give as much value as possible. Setup needs Node, PNPM, T-Mux, and the GitHub CLI, plus a dedicated agent GitHub account, so it's aimed at technical teams for now. For agencies, it's a free blueprint for the always-on teammate pattern that Claude Code and Codex are racing toward. Plus story. OpenAI will put ChatGPT and Codex on one shared token allowance starting in December. OpenAI updated its help center to say that starting in December, Plus, Pro, and business subscribers will draw on one allowance for both ChatGPT and Codex, measured in tokens. The change comes as OpenAI folds the agent features of work into the main ChatGPT app. Bigger models and harder tasks will use more of the allowance, while quick chats on instant models won't count at all. OpenAI says it expects most people won't hit a reset more often than they do today, but it hasn't shared token totals, plan figures, or an exact start date. It's also now capping how many paid ChatGPT accounts can run on one device and payment card, pointing heavy users to extra credits instead of a second subscription. For agencies running client work and coding agents on the same seats, it's time to track which tasks burn the most tokens before December. Three more headlines. First, Aramco closes its $2.6 billion deal for 49% of Saudi AI company Humane. Aramco now owns 49% of Humane, Saudi Arabia's national AI company, with the public investment fund keeping 51%, the two announced on Sunday. Aramco is paying about 9.9 billion Saudi reales in cash and handing over its own AI, computing, and digital assets. Hume is targeting 3 gigawatts of AI data center capacity by 2030. Second, semi-analysis launches GDPX, a tough agent benchmark where the best model scores 12%. GDPX includes 206 long agentic tasks across 11 high-value domains like chip design, cybersecurity, finance, and law. Claude Opus 5.5 leads at 12%, ahead of Gemini 4 Carbon at 10.9%, and GPT-6 Astra at 9.6%, and finance and legal tasks scored basically zero across every model. Carbon got close to the top at about a third of Opus's cost per task. Third, JPMorgan, Amazon, and other big brands are blocking outside AI agents like Metamuse. JPMorgan Chase, Amazon, Charles Schwab, Coles, and Neiman Marcus have moved to block personal agents like MetaMuse from acting for customers, while Walmart is opening its doors to them. An audit of 51 large U.S. retailers found 37 block shopping agents largely by accident through old bot rules, and none offers a way for a verified agent to complete a purchase. Tools of the Day. Beacon at Beacon.sh gives coding agents like Claude Code, Cursor, and Codex an open source memory layer that learns from past sessions so they stop repeating the same mistakes. Cedric at meat-cedric.com is an AI teammate for startups that works inside Slack, Discord, and Telegram, connects to thousands of tools, and can join your Zoom, Meet, or Teams calls to present and send follow-ups. Xmind2027 at xmind.com turns your mind map into an AI thinking partner with an Xmind agent that builds on your own documents and works alongside ChatGPT and Claude. Business AI tip of the day write attention rules before you turn on an always-on agent. DeepSeek built its super agent around one rule: take as little human attention as possible and give as much value as possible. That's the difference between an agent your client loves and one that floods their inbox. Before any agent runs around the clock for a client, decide exactly when it acts and when it speaks. So set up attention rules for one client agent this week. First, list who is allowed to give the agent work and ignore requests from anyone else. Write the triggers that let it act on its own, like a new lead, a failed payment, or a form submission. Define the only three reasons it may message a human, such as an approval, a blocker, or a finished result. Send everything else to a single daily digest instead of real-time pings. And log every action and message in sumo, AI, CRM, or GoHigh Level so you can review what it did and tighten the rules. Clients judge an agent by how much of their time it saves, not how much it does. Fewer, better interruptions are what make it feel like a real teammate. Pro tip track how many messages the agent sends per week. If that number climbs while results stay flat, your rules are too loose. That's your sumo growth AI Insights briefing. Write those attention rules this week, and I'll catch you on the next issue.