Your Dose of Reg.exe, Week {7}
Highlights this week include progress in durable agent execution, new infrastructure for AI-powered workflows, interpretability debates, and key model launches shaping the agent era.
Reg.exe is a global closed community of 260+ engineers, founders, and researchers interested in AI innovation, from San Francisco to Tokyo. Each week, we share the highlights of our discussions in a newsletter. If you’d like to join, write to join@welovesota.com
Ask for Help
🕵️♀️ Looking for Databricks/Snowflake users - Robert Hommes seeks companies struggling with augmenting human analyst workflows with AI, particularly those wasting time setting up notebooks for product analytics, fraud investigations, or regional deep dives. Contact: robert@moyai.ai or LinkedIn
Job Board
🕹 Gaming founding engineer opportunity - Former Balderton Capital GP and Google veteran building a gaming company are looking for a strong founding engineer passionate about gaming with AI/ML expertise (ideally computer vision). Role can be based in London, Paris, or hybrid. DM Margaux Wehr on Linkedin
Autonomous Agents
🛠 Durable agent execution with Temporal - Pydantic AI and Temporal released a durable agent setup similar to what OpenAI used for Codex and deep research. 📰 Learn about durable execution (🙏 Robert Hommes @ Moyai.ai)
🛎 PromptServer v1.0.0 released - Gabriel Olympie has built a simple backend server to serve Markdown prompts with full agentic and multimodal capabilities, production-ready and tested on thousands of queries per second. 📰 Explore PromptServer (🙏 Gabriel Olympie)
📖 YC starter pack on founder-led sales - Book recommendation for early-stage go-to-market strategies. 📰 Founding Sales handbook (🙏 Robert Hommes)
🎙 AI interpretability and OpenAI's road to AGI - Two significant interviews released discussing how AI models think and OpenAI's future plans. 🎞️ Anthropic's interpretability discussion and 🎞️ Greg Brockman on GPT-5 and beyond (🙏 Pierre Chapuis @ Finegrain)
Language Models
🧠 DeepSeek-V3.1 launched - First step toward the agent era featured hybrid inference with Think & Non-Think modes, faster thinking than DeepSeek-R1-0528, and stronger agent skills. The only model under 100B parameters in its performance tier. 🎞️ Announcement on X (🙏 Kevin Kuipers, Gabriel Olympie @ Reg.exe)
🧠 Reasoning in AI models discussion - Following the 🔬Chain-of-Thought Reasoning of LLMs a Mirage paper shared last week. Subbarao Kambhampati, who frequently discusses AI reasoning, is highly recommended to 📰 follow on X. Also (🙏 Marek Kalnik @ Theodo)
Programming
🛠 Claude Code alternatives emerging - New coding agent Amp gained traction among SF engineers as an alternative to Claude Code, promising autonomous reasoning and comprehensive code editing. 📰 Check out Amp (🙏 Robert Hommes)
🥷 AI disclosure requirements for open source - GitHub's Ghostty project now required contributors to disclose AI tool usage in their submissions, reflecting growing concerns about AI-generated code quality. 📰 View the pull request (🙏 Robert Hommes)
Infrastructure
⚙️ GitHub's AI-powered CI/CD revolution: Continuous AI and Agentic Workflows - GitHub Next introduced two groundbreaking projects exploring how AI agents can transform software development workflows.
Continuous AI (CAI) is a new category of AI-powered automation designed to support software collaboration, aiming to become a default layer in dev workflows like CI/CD. Tasks included auto-generating documentation, summarizing issues, suggesting better tests, and creating creative outputs from PR activity. 📰 Read about Continuous AI
Agentic Workflows is built on GitHub Actions and allows developers to describe intent in natural language (e.g., "scan coverage and add tests") and compile it into version-controlled workflows powered by LLMs like Claude or Codex. Concrete examples included auto-updating docs after code changes, personalized onboarding guides for new devs, accessibility checks with auto-raised PRs, security patches triggered by advisories, and continuous triage and summarization of issues. 📰 Agentic Workflows
The projects also explored guardrails to manage risks like prompt injection or model drift, aiming for human review and traceability. (🙏 Thierry Abalea @ Shipfox)
Cyber
⛓ SHA3 sponge function explained - Deep dive into SHA3 as the backup for SHA2, understanding how the sponge function worked and why it's ready if vulnerabilities are found in SHA2. 🎞️ Computerphile video (🙏 Kevin Kuipers @ Reg.exe)
Events
🇳🇱 PyData fintech MCP event - August 26 at Mollie Amsterdam office featuring MCP server live-coding and feature selection data hacks. Register for the meetup (🙏 Robert Hommes)
Geographic Hubs
🇺🇸 San Francisco
🏐 Community meetups and connections - Robert Hommes (Moyai.ai) and Benjamin Trom (Mistral AI) enjoyed some volleyball activities. Philippe Mizrahi (Linkup) was at the AWS Startup Loft.
🇯🇵 Tokyo
🍣 Sushi meetup with 2501.ai - Alex Pereira and Alex Zhuk from 2501.ai enjoyed traditional sushi in Tokyo with Kevin Kuipers.
