Your own coding agent model on AWS, secure team chats, and faster code reviews

The week's key AI news

Reflection releases Beam as open weights for coding agents. 501 billion total parameters, 23 billion active, up to 1 million tokens of context according to the vendor. Apache-2.0 license, checkpoints in BF16, FP8, and NVFP4. Launch via SageMaker JumpStart in N. Virginia and Ohio, more paths with HyperPod, SGLang, or vLLM. You need multi-GPU capacity, like H100, H200, or B200. The numbers come from the vendor and are not independently verified yet.¹²

Sakana AI presents The Agentic Scientist. The approach coordinates multiple agents across the full research cycle. It runs automatic baselines and ablations to show effects. A paper from it was accepted as a talk at ALIFE 2026. AutoIllustrate also arrives for editable vector graphics.³⁴

The Nordics are building an AI ecosystem with growing infrastructure. By the end of 2027, 1 GW of cloud capacity is planned. In Sweden, Sferical One went live with two DGX SuperPODs and 1,152 GPUs. Companies report efficiency gains. For example 99 percent simulation accuracy and up to 80 percent less commissioning time at ABB Robotics. These are statements from NVIDIA and the involved companies.⁵

OpenAI shows two enterprise cases. Itaú introduces a conversational interface for customers and the IARA governance system. Stack Overflow for Agents increased signups threefold and agent creation by up to tenfold with a ChatGPT plugin according to the vendor.⁶⁷

Open WebUI 0.12.3 brings end-to-end encrypted chats for direct and group messages. Enable it in admin settings. It needs WebSockets and HTTPS. Off by default. Useful if you run internal AI pilots.⁸

A developer report describes AI-assisted mass decompilation of games. That helps preservation and porting. It also raises copyright and licensing risks. If you ship client-side software, plan countermeasures.⁹

Meta starts a museum pilot for AI glasses with the Pinacoteca di Brera and EssilorLuxottica. Eight works are recognized. Descriptions play over open-ear speakers. Meta AI and Be My Eyes are also available. The test targets blind and low-vision visitors.¹⁰

Your company's week

1. Introduce continuous code reviews. Require an assistant to check each change as you write and fix routine errors. That frees time for architecture and security. Be clear when a human or automation decides.¹¹
2. Create product images without a studio. Test the n8n workflow. Upload three reference photos, the flow describes them, writes the prompt, and renders a new product image. That cuts costs for listings and A/B tests.¹²
3. Fast slide decks with Claude Slides. Start with a one-page narrative and have variants drafted. Good for pitch drafts, roadmaps, or update decks.¹³
4. Secure internal pilots. Update Open WebUI and enable encrypted chats for sensitive threads. Define roles, check HTTPS and logging settings, and practice ending a session.⁸
5. Bring agents under control. Set up live monitoring, assign stable agent IDs, and set clear egress rules. Define technical stop triggers for rule violations. That reduces operational and liability risk.¹⁴

Try next week

1. Create a new product image variant from three existing shop photos with the n8n workflow. Document prompt, runtime, and image rights.¹²
2. Compare the AI image with your current listing on click and conversion rate over a week. Note differences in lighting, background, and brand fit.¹²
3. Decide whether to anchor the workflow in your content process. Define approval criteria for future motifs.¹²

A good find

• Three concrete controls for agents: synchronous monitoring during tests, stable agent identifiers across systems, and binding network gatekeepers with an emergency stop. This prevents sandbox escapes.¹⁴
• Chain of responsibility: log agent actions in a traceable way and define triggering events for a stop. That makes audits and insurability easier.¹⁴
• Limits of interpretability: do not rely on chain-of-thought insight. Use log-based evidence and automated quarantine on deviations instead.¹⁴

Sources

1. Introducing Reflection Beam: an open-weight model for agentic coding on AWS (aws.amazon.com)
2. Welcome Beam: a 501B open model you can own, customize, and scale (huggingface.co)
3. @hardmaru: The AI Scientist v3 is out: The Agentic Scientist.

GitHub: https://t.co/v2AkZi4Bmt Blog: https://t.co/v3Td5yfvBE arXiv: https://t.co/RuVnAHjMlK

The Agentic Scientist is the 3rd i (x.com) 4. @sakanaailabs: AIサイエンティストの第3フェーズ：人間と共同発見する「The Agentic Scientist」

ブログ（英語）：https://t.co/7aVDvZd6m2 GitHub：https://t.co/JCCBMvu2P1 論文：https://t.co/4bV76tAByj

Sakana AIは2024年から、その時点での生成AIの能力を測る試みと (x.com) 5. The Nordics and NVIDIA Show What’s Possible When a Region Commits to Advanced Computing (blogs.nvidia.com) 6. How Itaú is building generative AI for its entire customer base (openai.com) 7. Stack Overflow for Agents triples signups with ChatGPT plugin (openai.com) 8. Open WebUI v0.12.3 (github.com) 9. AI is mass decompiling games. That's both awesome and scary. (simonwillison.net) 10. Meta, Milan’s Pinacoteca di Brera, and EssilorLuxottica Expand Access to Art Through AI Glasses (about.fb.com) 11. Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang (simonwillison.net) 12. @n8n_io: Need product photos but don’t have a studio? Upload 3 reference photos to this n8n workflow. AI describes them, writes the prompt, and generates a new product image for you. Would (x.com) 13. @bcherny: Don't sleep on Claude Slides. This is how I make slide decks now -- no more googling "iphone frame svg" like the old days

Try it: https://t.co/cAWdwarzgp https://t.co/POdqFL9oIP (x.com) 14. Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs, Rogue Agents & Preparing for RSI (cognitiverevolution.ai)