Databricks released Agent Bricks, a new product that helps enterprises develop SOTA domain-specific agents.

Agent Learning from Human Feedback (ALHF) is a new paradigm where agents learn directly from minimal natural language feedback, not just labels or numeric rewards.

Google released a brilliant research, a new active learning method for curating high-quality data that reduces training data requirements for fine-tuning LLMs by orders of magnitude.

You get the same or better model quality with 250 to 450 expert labels instead of 100K, thanks to focusing expert effort only on confusing boundary cases.

The team uses an LLM as a scout to sweep a huge pool of ads, then asks experts to judge only the handful of examples that truly confuse the model.

Those expert decisions train the next model iteration, and the loop repeats until model–expert agreement stops improving. It is classic active learning, but adapted to large, noisy, imbalanced traffic like ads, where only about 1% of items are actually problematic.

Big AI security issue: ChatGPT vulnerability in Connectors can leak your API keys + "memory" with 0 clicks

AgentFlayer, demoed at BlackHat, shows an injected prompt in a doc can force an image to render that exfiltrates data through a malicious URL.

Tencent AI Lab introduced R-Zero. This framework enabling LLMs to self-evolve their reasoning capabilities

From zero human-curated data, through an autonomous Challenger-Solver loop.

R-Zero learns from scratch, with the Challenger proposing tasks at the edge of the Solver's ability.

This co-evolution boosts Qwen3-4B-Base by +6.49 on math and +7.54 on general reasoning.

GitHub.

Renmin University of China and Huawei presented comprehensive survey of memory mechanisms in LLM-based agents:

• What memory is & why it matters
• How to design & evaluate it
• Key applications & use cases
• Limitations & future directions

A roadmap for building smarter, longer-lived AI agents.

GitHub.

The top AI agents by revenue @alwebbci

The AI agent market expected to 2x+ this year ($5B to $13B). 50% of top 20 were founded in the last 3 years.
Customer service AI agents command 127x revenue multiples vs. 52x average.

Alibaba introduced Memp, a new framework that gives LLM agents learnable, updatable procedural memory.

This leads to steadily higher success rates and greater efficiency on complex tasks.

Memp distills past agent trajectories into both fine-grained instructions and high-level abstractions, continuously improving with new experience. It's even transferable to weaker models.

Circle has announced the launch of Arc, an open Layer-1 blockchain designed to provide enterprise-grade infrastructure for stablecoin payments, foreign exchange, and capital markets applications.

The network is EVM-compatible and uses USDC as its native gas token. Arc is expected to launch its public testnet later this fall.

Microsoft has released a 30-page guide on #AIAgent governance to help secure and manage agents in #Microsoft365 environments

Anthropic: Claude Sonnet 4 now supports 1 million tokens of context on the Anthropic API—a 5x increase.

Process over 75,000 lines of code or hundreds of documents in a single request.

Long context support is in public beta for API users with Tier 4 and custom rate limits.

Broader availability will be rolling out over the coming weeks. Available in Amazon Bedrock, and coming soon to Google Cloud's Vertex AI.

Microsoft introduced Dion is a new AI model optimization method that boosts scalability and performance over existing leading methods by orthonormalizing only a top rank subset of singular vectors, enabling more efficient training of large models such as LLaMA-3 with reduced overhead.

Orthonormal updates appear to roughly double transformer training convergence with Dion paving tractability at the largest scale.

Code.
Paper.

Matrix-Game 2.0 — the first open-source, real-time, long-sequence interactive world model

Last week, DeepMind's Genie 3 shook the AI world with real-time interactive world models.

But... it wasn't open-sourced.

Matrix-Game 2.0 is Skywork's next-gen interactive world model:

- Real-time: 25FPS generation
- Long-sequence: Minutes of continuous video
- Interactive: Move, rotate, explore
- Multi-scene: City, wild, TempleRun, GTA.

It's the foundation for:

- Game engines
- Embodied AI
- Virtual humans
- Spatial intelligence.

The Tech Stack:

- Data: 1,350 hrs of interactive videos from Unreal Engine + GTA5
- Control: Frame-level keyboard & mouse input
- Model: 1.3B autoregressive diffusion with action control
- Speed: Single GPU → 25FPS
- 3D Causal VAE for space-time compression
- Diffusion Transformer with action conditioning
- KV-Cache for infinite video generation
- DMD training to avoid error accumulation

Now you can run and benchmark evolutionary coding agents on 100+ algorithm optimization tasks from algotune.io

Google is rolling out their version of memory for Gemini today. It is called 'personal context.'

If you want to disable this, toggle off Personal Context in settings.

This works for 2.5 Pro only, not Flash.

It will be interesting to see what the effect of Gemini's monster context window will have on implementation.

Anthropic acquired the co-founders and most of the team of Humanloop, the startup behind Hooli, a platform for prompt management, LLM evaluation, and observability

The move marks a major push from Anthropic to strengthen its AI safety strategy.

ByteDance & Tsinghua University unveiled ASearcher

Agentic search by enabling long-horizon reasoning with large-scale asynchronous RL.

Goes beyond typical turn limits for complex, knowledge-intensive tasks.

Achieves SOTA performance, with significant gains of up to +46.7% on xBench and GAIA after RL training.

Models & data.