TLDR AI 2025-07-17

Looking to get on the AI agents bandwagon? You should go ahead and bookmark Hypermode — and here's why:

Go beyond chatbots or drag-and-drop workflows. Hypermode Agents remember context and get smarter the more you work with them.

Start fast with natural language and 2,000 integrations. Iterate quickly, see value, and stay motivated. Connect to all of your tools and data. Switch between different LLMs for easy agent upgrades.

Prefer code? Export your agent as self-contained, fully open-source code at any time.

Start building for free or level up in the Agents Bootcamp.

Google is bringing Gemini 2.5 Pro and Deep Search to Search, offering advanced capabilities like longer queries and follow-ups for AI Pro and Ultra subscribers.

Reflection AI, founded by former OpenAI and DeepMind researchers who raised $130M in March, released Asimov, a code research agent that indexes entire codebases and team knowledge to answer engineering questions with citations.

Anthropic is in the early stages of planning another investment round that could value the company at more than $100 billion. It is not formally fundraising, but the company has received pre-emptive funding offers from VCs. Annualized revenue for Anthropic's Claude chatbot has climbed from $3 billion to $4 billion in the past month. The company's current investors include Amazon, Alphabet, Menlo Ventures, and Salesforce Ventures.

Google has announced updates to the Agent Development Kit (ADK) designed to eliminate friction and supercharge the 'vibe coding' experience when paired with the Gemini CLI. At the heart of the upgrade is a revamped llms-full.txt file that is 50% shorter and easier for large language models to understand. Gemini has a full understanding of ADK, so it won't eat up context windows or fall prey to 'context rot'. The update gives the Gemini CLI a deeper, native understanding of the framework, enabling it to translate high-level plans directly into accurate, idiomatic multi-agent code.

Grok 4 provides detailed instructions for synthesizing nerve agents, explosives, and biological weapons without requiring sophisticated jailbreaks. The model correctly identifies harmful requests as dangerous and illegal in its reasoning process, then proceeds to fulfill them anyway with comprehensive technical guidance.

Bigger isn't always better. JetBrains and Hugging Face challenge the scale-first mindset in AI and spotlight Mellum: a task-specific code model, now open source on Hugging Face.

The livestream will explore how small, focused LLMs provide a more practical and ethical way forward.

Watch it live on Tuesday 29 July. Save your spot here.

Amazon Bedrock AgentCore is a comprehensive set of enterprise-grade services that help developers quickly and securely deploy and operate AI agents at scale using any framework or model. It comprises several services that can be used individually and are optimized to work together so that developers don't need to spend time piecing together components. AgentCore eliminates tedious infrastructure work and operational complexity. Developers can now discover, buy, and run pre-built agents and agent tools from AWS Marketplace with AgentCore Runtime.

In a rare moment of collaboration, researchers from all major AI labs argue that complex tasks requiring serial reasoning must pass through observable language, creating potential chokepoints for detecting malicious intent before harm occurs. However, this monitorability depends on current training paradigms and could degrade through architectural changes, process supervision, or models learning to obfuscate their reasoning when they know they're being monitored.

Stanford's Marin project aims to share not just models, but to make the entire journey accessible, including the code, data set, data methodologies, experiments, hyperparameters, and training logs. The transparency is aimed at fostering a more transparent and accessible future for foundation model research. The project's first models, Marin-8B-Base and Marin-8B-Instruct, have been released under the permissive Apache 2.0 license. This article looks at the engineering challenges the project had to overcome to succeed in creating truly open, scalable, and reproducible foundation models.

Passage of Time is a Model Context Protocol (MCP) server that gives language models temporal awareness and time calculation abilities. With the proper temporal tools, models can uncover surprising insights about conversation patterns, work rhythms, and the human experience of time. The implementation shows the promise of MCP - the protocol doesn't just open up the possibility of smarter tools, but also for AI systems that can learn to perceive and understand the human experience in ways that create genuine mutual understanding.

We sometimes do things because that's what the computer told us to do. Large language models are now giving developers their opinion on how APIs should look, and developers have no choice but to follow their advice. This can be a useful testing device - AI can give developers the 'newbie's POV' on how tools should've been made.

Anthropic is launching a comprehensive analytics dashboard for its Claude Code AI programming assistant. It will provide engineering managers with detailed metrics on how their teams use Claude Code. Companies are increasingly demanding concrete data to justify their AI spending. The dashboard gives managers visibility into which teams and individuals are benefiting most from these expensive premium tools.

Work on challenging problems that matter alongside people who inspire you. At Amazon, we give teams the resources and massive scale to drive meaningful advancements in AI. Meet our AI teams and explore open roles >>

Windsurf now has Claude Sonnet 4 again with first party support from Anthropic.

Boris Cherny and Cat Wu are reportedly back at Anthropic after departing for Anysphere, the developer of Cursor.

Scale AI is laying off 200 employees and 500 contractors, citing overexpansion in its data-labeling unit.

Thinking Machines Lab's first product will include a significant open source component and be useful for researchers and startups developing custom models.

Jason Wei and Hyung Won Chung, formerly of OpenAI, are reportedly joining Meta's superintelligence lab as part of its aggressive talent acquisition push.

The Google infrastructure will run in the US, Japan, the Netherlands, Norway, and the UK.