01AgentsInternals5 sources agree
Researchers at the Beijing Academy of Artificial Intelligence developed AREX, a research agent that uses repeated cycles of gathering evidence and reflecting on provisional answers to improve its performance. The agent outperformed competing approaches on several benchmarks, with fine-tuning and targeted training on critical decision points contributing to its success. The results demonstrate the effectiveness of iterative harnesses and fine-tuning in improving research agent performance.
Provenance — who else covered this
02ModelsProduct5 sources agree
Drex 1.5 has been released with a 128k context window, allowing it to process whole contracts and large candidate pools, and reports 93.4% accuracy, also ranking #1 on Decision Index 0.2.1.
Provenance — who else covered this
03AgentsProduct3 sources agree
Oracle AI Agent Memory implements short-term and long-term memory for agents, allowing them to store and retrieve state across sessions, and outperformed flat history in an 80-turn evaluation. The system's memory architecture is part of a larger agent harness that includes an orchestration loop, tools, state persistence, and guardrails.
Provenance — who else covered this
04AI securityInternals3 sources agree
Claude's API was found to have a misconfiguration that leaks user data, including emails and usernames, while GLM had a token forgery method discovered, and a deeply embedded SSRF was found, though it's unclear if it's an n-day or 0-day vulnerability.
Provenance — who else covered this
05AgentsProduct2 sources agree
MCP Apps is a standard for running interactive apps inside AI chats, already used by Booking.com, Figma, and Canva, allowing developers to integrate their products into chat platforms like ChatGPT and Claude, and a developer builds an MCP app from start to finish using the open-source Skybridge framework.
Provenance — who else covered this
06AgentsProductsingle source
Pi has released a Session Manager API for programmatic creation, persistence, and tree navigation of sessions, which are stored as JSONL files with a tree structure via id/parentId fields. The API supports various session entry types, including messages, compactions, branch summaries, and custom entries.
Provenance — who else covered this
07ModelsProduct8 sources agree
Cloudflare has released an alternative to Jev, hosted on Hugging Face, under the Apache 2.0 license.
Provenance — who else covered this
08AgentsProduct4 sources agree
CopilotKit has released OpenDots, an open-source version of OpenAI Dots that can be self-hosted and integrated with any agent harness, available on mobile and web platforms.
Provenance — who else covered this
09ModelsProduct3 sources agree
A state-of-the-art multimodal Decision Model called pplx-decider-27b is being open sourced and offered through a new Decisions API, initially priced at 4 cents per million input tokens, with plans for future price reductions.
Provenance — who else covered this
10ModelsProduct2 sources agree
Strands introduced Strands Decider 2B, a small open-source decision model, and Cloudflare announced Clef, an open-source decision model with a new RL fine-tuning platform, both aiming to improve decision-making capabilities in AI systems.
Provenance — who else covered this
11ModelsProductsingle source
Treg AI has open-sourced their Jev implementation on GitHub, which they claim is 10x cheaper, 18x faster, and 30% more accurate than GPT6-luna in their tests, used for tasks such as classifying web traffic and smart onboarding.
Provenance — who else covered this
12AI securityProduct5 sources agree
A security engineer discusses the ineffectiveness of annual vulnerability cycles and the need for security teams to own and fix issues, rather than relying on vendors or MSSPs. They also emphasize the importance of rehearsing incident response and providing actionable findings to infrastructure teams. The engineer criticizes the 'build versus buy' approach and advocates for a more honest assessment of what can be built in-house.
Provenance — who else covered this
13AgentsInternals3 sources agree
A Google paper introduces a pattern where a verifier starts from a clean context to validate the explorer's reasoning, allowing for the exploration of wild ideas while discarding incorrect proofs, similar to how deepagents subagents use separate contexts.
Provenance — who else covered this
14AgentsProduct2 sources agree
The Agents API team has introduced several new features, including the ability to spin up a browser and agent with one API call, Bedrock Managed Agents, and environment portability, as well as improvements to the dashboard and overall reliability.
Provenance — who else covered this
15AI securityProductsingle source
GLM, Kimi, and Claude were pitted against an authorized target in a pentest, with each model taking a different approach to vulnerability discovery, including dynamic web pentesting, code inspection, and patch differentials for n-day CVEs. The results showed varying degrees of success in identifying info leaks and potential vulnerabilities.
Provenance — who else covered this
16CodingProduct2 sources agree
Anthropic's Claude AI now allows customization through mods, which can be shared as plugins, enabling personalized experiences. This feature enables users to tailor Claude to their individual workflows and preferences.
Provenance — who else covered this
17AgentsProduct2 sources agree
Multiple companies including Perplexity, Cloudflare, and OpenAI have released decision models, with OpenAI's Decisions API currently in limited preview, following the release of JEV in September.
Provenance — who else covered this
18AgentsProductsingle source
LlamaParse's agentic parsing tier can understand and extract data from the BEA's Q2 GDP-by-state infographic, converting it into a 52-row table. The original data is only available in infographic format without a text layer.
Provenance — who else covered this
19CodingProductsingle source
A developer built an example using Claude code mods to access GitHub PRs and issues by typing #github, demonstrating the feature's potential for extensibility.
Provenance — who else covered this
20AgentsProduct6 sources agree
Nhu Hoang explores graph engineering ideas and their application to real workflows, discussing the changes when an AI agent is designed as a graph instead of a simple loop.
Provenance — who else covered this
21AgentsProduct2 sources agree
Anthropic has renamed the '/claude-api prompt-audit' command to '/checkup prompt-audit' to better reflect its functionality, which checks for unnecessary model instructions in CLAUDE.md, skills, and agents.
Provenance — who else covered this
22CodingProduct2 sources agree
A new plugin for Claude Code, called You Should Know, provides important information that users might have missed.
Provenance — who else covered this