On — Nov 12

AWS releases AWS Emulator, a free local emulator of AWS services for building and testing apps on your own machine or in CI, with no deployments and no AWS account. It is generally available, and AWS says every line of its code was written by autonomous agents running on Amazon Bedrock AgentCore.

Why it spiked. Hundreds of agents built it with almost no human-written code, and it gives coding agents a safe place to test against AWS.

• @SwamiSivasubram Says AWS Emulator gives coding agents a local copy of AWS to test against because agents now write code faster than teams can check it in production.
• @awsdevelopers Announces general availability for local and CI testing with no deployments and no AWS account.
• @AWSNewsroom States that hundreds of agents, organized into three teams, wrote all of the code on Bedrock AgentCore.

Google launches the Gemini Decisions API, powered by a new Gemini Decision 1 model in Google AI Studio and the Gemini API. Instead of generating text, it scores closed-set choices for up to 1,000 questions in one pass across text, images, audio, video, and PDFs, with a 256K context window at $0.10 per 1M input tokens and no additional cost.

Why it spiked. Google says it uses the model internally for agent routing and grading, and it charges only for input tokens.

• @OfficialLoganK Lists $0.10 per 1M input tokens with no additional cost, native image, audio, video, and PDF input, 256K context, and 1K questions, free to try in the API.
• @sundarpichai Describes it as built for closed-set choices, scoring up to 1,000 questions in a single pass.
• @GoogleAIStudio Says it returns calibrated decisions and is available in AI Studio and the Gemini API.

Google Antigravity agents can now control the built-in browser for online research, localhost testing, and checking their own work. Users choose what the agent logs into, downloads wait for approval, and the agent can run several browser tasks at once. It ships in v2.31.0 for Free, Pro, and Ultra, with Enterprise coming later behind admin controls.

Why it spiked. It is the third Antigravity launch in three days, and the team teases another for Thursday.

• @antigravity Announces the rollout with user-controlled logins, approval for downloads, and parallel browser tasks.
• @kevinhou22 Says to ask the agent to use the browser or invoke the /browser skill after updating to v2.31.0.
• @_anshulr Lists the week so far: knowledge work layout Monday, Carbon and /swarm Tuesday, browser use Wednesday, and a teaser for Thursday.

Managed Agents on the Gemini API now support self-hosted environments. Google keeps running the model and agent loop, while code execution and file operations run in a sandbox the developer controls.

Why it spiked. Sandbox providers announced same-day support, which lets teams keep data and network rules on their own infrastructure.

• @GoogleAIStudio Announces bring-your-own-sandbox through self-hosted environments.
• @flydotio Says every command and file can stay in a microVM in your own Fly org while Google handles the model and agent loop.
• @rakyll Says Managed Agents now supports Agent Substrate as a self-hosted environment for running operations on your own clusters.

Runway releases Praxis-1, a 27B world action model that turns video pretraining into control for real robots. Weights are open under Apache 2.0, and Runway says it ranks first on NVIDIA's RoboLab-120 leaderboard, ahead of open-weight VLAs including π0.5 and GR00T N1.6.

Why it spiked. It is the top open robotics model on that leaderboard, and LeRobot and NVIDIA added support on day one.

• @runwayml Releases Praxis-1 with open weights under Apache 2.0, ready to run and fine-tune on your own hardware.
• @LeRobotHF Adds it to LeRobot and says one checkpoint adapts to a new robot with light fine-tuning.
• @NVIDIARobotics Notes the 27B size and support from data center GPUs down to Jetson, with an NVFP4 download.

OpenAI researchers publish work on how chain-of-thought monitorability changes over training. They report that no-CoT capability and CoT controllability both rise as models scale, so models may become harder to monitor by reading their reasoning.

Why it spiked. Safety researchers read it as a warning about a main oversight method that labs depend on.

• @jeffreygwang Opens a 10-post thread on the research and calls the findings unsurprising but sobering.
• @_aidenlow Says early results indicate models may become increasingly difficult to monitor through their CoT.
• @omarsar0 Recommends the paper as a setup for studying how and where this effect emerges.