Subscribe, if this was forwarded to you

Hey there,

This week’s edition is packed with groundbreaking advancements and practical tools at the frontier of AI and automation. We spotlight OpenAI’s GPT-5.6 Sol Ultra for its historic proof of the Cycle Double Cover Conjecture, and take a deep dive into agentic AI with the newly released “Hitchhiker’s Guide to Agentic AI.” You’ll also discover Browser Use, an open-source agent redefining browser automation, and get actionable strategies for slashing AI costs with Anthropic’s “Plan Big, Execute Small” pattern. Plus, don’t miss the latest on ChatGPT Work, GPT-Live, and cutting-edge open-source security and math models.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Agentic AI foundations and deployment guide

🛠️ Tool Spotlight: Browser Use: open-source browser automation agent

📰 Top News: GPT-5.6 Sol Ultra proves CDC conjecture

💡 Tip: Plan Big, Execute Small for AI savings

Let's get started!

Tutorial of the week

Mastering Agentic AI: From Foundations to Deployment

A new, freely available book—The Hitchhiker's Guide to Agentic AI: From Foundations to Systems—offers a comprehensive roadmap for building advanced, autonomous AI agents. This resource is a must-read for developers and researchers eager to understand the entire stack of agentic AI, from core theory to practical system implementation.

• Covers the full pipeline: transformer architectures, GPU systems, training/fine-tuning (SFT, LoRA, MoE), model compression, and inference optimization.

• Deep dives into alignment and reasoning: RLHF, PPO, DPO, reward modeling, and advanced reasoning techniques like chain-of-thought and test-time scaling.

• Explores agentic AI in practice: agentic training, retrieval-augmented generation (RAG), memory systems, agent harness design, and context management.

• Examines inter-agent coordination: Model Context Protocol (MCP), agent skills/tool use, A2A communication, and multi-agent architectures.

• Practical guidance throughout: each chapter blends rigorous theory with implementation tips, code examples, and references to primary literature.

Whether you're building your first AI agent or architecting complex multi-agent systems, this book provides the technical depth and practical insights needed to succeed. Dive in to accelerate your journey in agentic AI.

Tool of the week

Browser Use — Open-source AI agent for browser automation

Browser Use is an open-source Python library and agent that empowers AI models to interact with web browsers just like a human—navigating pages, clicking buttons, filling forms, and extracting data. It’s designed for developers who want to automate complex web tasks, build robust web scrapers, or integrate browser actions into AI workflows without relying on expensive APIs or proprietary solutions.

• Human-like browser control: Automate any web task—form filling, data extraction, QA testing, or site navigation—by simply describing the task in natural language.

• Flexible LLM integration: Works with OpenAI, Anthropic, Google, or its own optimized ChatBrowserUse models; bring your own API keys or use the free open-source agent.

• Scalable and production-ready: Choose between running locally for full control or leveraging the hosted cloud agent for advanced features like proxy rotation, stealth mode, captcha solving, and 1000+ integrations (Gmail, Slack, Notion, etc.).

• Customizable and extensible: Deep code-level integration lets you add custom tools, automate at scale, and embed browser agents into your own products.

• Benchmark leader: #1 on the Odysseys leaderboard (87.4% average on 200 real-world web tasks), outperforming agents from OpenAI, Anthropic, Google, and Microsoft.

Browser Use is rapidly gaining traction in the AI automation and web scraping community, with an active open-source ecosystem and robust documentation. Try it for free and supercharge your browser automation projects.

Top News of the week

OpenAI’s GPT-5.6 Breaks New Ground: Proves Cycle Double Cover Conjecture

OpenAI’s GPT-5.6 Sol Ultra has achieved a landmark breakthrough by producing a proof for the Cycle Double Cover Conjecture, a longstanding unsolved problem in graph theory. Remarkably, the proof was generated in under one hour using 64 parallel subagents, showcasing the immense potential of multi-agent AI systems in tackling complex, open-ended research challenges. This accomplishment signals a new era for AI-assisted mathematical discovery and advanced scientific problem-solving.

The proof, along with the prompt and a formal Lean verification, is now publicly available on GitHub. The formalization leverages the Jaeger–Kilpatrick eight-flow theorem, constructing a nowhere-zero flow and converting it into a cycle double cover for finite loopless bridgeless multigraphs. The Lean project is pinned to version 4.31.0 and uses a specific Mathlib revision, ensuring reproducibility and transparency for researchers wishing to audit or extend the work. While the proof has not yet undergone peer review, its kernel-checked formalization provides a strong foundation for future validation.

This milestone demonstrates the power of parallelism and collaboration among AI agents in solving deep mathematical problems, sparking excitement and debate within the research community about the future of AI-driven science.

🔗 Read the official CDC Lean formalization and proof

OpenAI Launches ChatGPT Work for Autonomous Workflow Automation

OpenAI has introduced ChatGPT Work, a new agent mode powered by GPT-5.6 that can autonomously execute complex workflows across apps like Google Drive, Slack, and Salesforce. ChatGPT Work breaks down high-level goals into actionable steps, producing finished documents, slide decks, spreadsheets, or web apps with minimal user intervention. The feature is rolling out to Pro, Enterprise, and Edu users, with Plus and Business access coming soon. This marks a major leap in workplace productivity, enabling teams to delegate repetitive tasks and focus on higher-value work.

🔗 Read more at OpenAI

T3MP3ST: Open-Source AI Red Teaming Tool Democratizes Security Testing

Security researchers have released T3MP3ST, a free and open-source framework that transforms AI coding agents like Codex and Claude Code into automated security testers. T3MP3ST can red-team web apps, APIs, smart contracts, and IoT systems, achieving impressive results on industry benchmarks and real-world vulnerabilities. With 35+ built-in tools and support for local or cloud models, it enables developers and organizations to conduct advanced security assessments without specialized hardware or costly licenses.

🔗 Check out T3MP3ST on GitHub

Mistral Releases Leanstral 1.5: Open-Source Model for Advanced Math Reasoning

Mistral AI has launched Leanstral 1.5, an open-source code agent model designed for Lean 4, a proof assistant used in complex mathematical and software verification tasks. The model boasts 119B parameters and supports multimodal input, offering high performance and cost efficiency. Leanstral 1.5 is capable of tackling long-range, multi-hour reasoning tasks and is available for free use via Mistral’s API or local deployment. This release further democratizes access to advanced mathematical problem-solving tools.

🔗 Learn more on Hugging Face

OpenAI Unveils GPT-Live: Real-Time, Full-Duplex Voice AI

OpenAI has released GPT-Live, a next-generation voice model enabling full-duplex communication—allowing ChatGPT to listen and speak simultaneously. This results in more natural, conversational interactions, with the AI able to acknowledge, interrupt, or pause as needed. GPT-Live also delegates complex tasks to advanced models in the background, ensuring smarter and faster responses. The update is rolling out globally for both free and paid ChatGPT users, with API access on the horizon. This advancement sets a new standard for human-AI voice interaction.

🔗 Explore GPT-Live at OpenAI

Tip of the week

Cut AI Costs with "Plan Big, Execute Small" Agent Patterns

Running production-scale AI pipelines with frontier models can get expensive fast. Anthropic’s “Plan Big, Execute Small” orchestration pattern lets you achieve nearly top-tier performance while slashing costs—by using cheaper submodels for most tasks and only calling the most powerful model when absolutely necessary.

• What it is: This workflow delegates routine or lower-complexity tasks to cost-effective submodels (like Claude Haiku), reserving the frontier model (like Claude Opus) for critical steps that require maximum capability.

• How to apply: Use the Claude Managed Agents API to set up a pipeline where sub-agents handle planning, retrieval, or summarization, then escalate to the main agent for final synthesis or complex reasoning. Example:

• Why it’s useful: This approach can deliver up to 96% of frontier model performance at nearly half the cost, making advanced AI more accessible and scalable.

• Key benefits: Native support for caching, context handling, and seamless orchestration in Claude Managed Agents.

Try this pattern when you need to optimize for both performance and budget in your AI workflows.

Explore the Claude Cookbooks for orchestration examples

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help