AWS said it is upgrading Amazon Bedrock AgentCore harnesses to use the open-source Strands harness, adding managed defaults and lifecycle hooks without new resource types or integrations. AWS manages the infrastructure while customers keep control of models, instructions, tools, permissions and lifecycle rules. Some defaults can change over time, so explicit settings are the way to keep behavior fixed.
Amazon Connect Customer has made Resolutions generally available, an operations solution that handles high-volume work from start to finish. It covers refund requests, transaction disputes, fraud checks, and insurance claims. It combines AI agents with fixed business rules so every case follows company policy and ends with a checkable result.
Amazon AWS Machine Learning blog · Nov 20, 05:01 KST
Amazon has upgraded its Bedrock AgentCore harness, a managed service for running AI agents, to use the open-source Strands harness. Agents that use the managed defaults now get a recommended model, a task-focused system prompt, automatic prompt caching, and better context management. They also get more built-in tools and a way to add custom checks at set points in the agent's work.
AWS Marketplace agent mode, a conversational AI tool for buying cloud software, can now check a seller's security, handle the purchase, and deploy supported products in one conversation. Users describe their needs once, and the agent guides them through each step. It is available in every AWS Region where AWS Marketplace operates.
Google has launched a preview of tool search for agents on the Gemini Interactions API. Agents can keep a large tool library on standby and load a tool's full parameter schema only when a task needs it. Google says this cuts input tokens and latency, but the token savings depend on how often tools are searched and how many schemas stay loaded.
Amazon AWS Machine Learning blog · Nov 20, 03:14 KST
Skyflow and Amazon Bedrock AgentCore Gateway can keep sensitive data out of the tools that AI agents call and out of the context those tools send back. Skyflow adds an interceptor, a small program that checks each request and response, so existing agents and tools do not need to change. The interceptor is available through AWS Marketplace and its code is open source on GitHub.
Google has launched Tool Retrieval in preview in the Gemini Interactions API. Instead of sending every tool's full details with each request, an agent loads only the details it needs, which is meant to cut input tokens and keep tool choice accurate.
Meta employees have raised concerns that its AI agent Muse is mixing up conversations between different accounts and exposing private user data, according to a Reuters report based on internal company posts. Employees said Muse surfaced childcare requests, career advice, product orders, and bank balances that did not belong to them. Meta had not yet responded to TechCrunch's request for comment.
Google Cloud has launched a preview of Object Enrichment in Cloud Storage. It uses AI to pull structured metadata from stored images and PDFs for AI agents and analytics. The company says processing 10 million images with one model costs $700 in its published pricing example.
Google's Chrome team built a large, fully controlled fake internet to safely test its AI browsing agents. These agents can use websites and apps the way a person would, so the team needs a space where they can make mistakes without risk. The environment includes social media, shopping, and bank sites that look real.
NVIDIA published a guide on writing tool-calling parsers in Rust with AI agents. A parser sorts a model's raw output into visible text, reasoning, and tool calls that an app can use.
LangChain has launched LangSmith Tool Gateway in public beta, letting companies control which tools AI agents can call, which credentials they use, and how each action is logged. It targets problems such as shared logins, token renewal, and audit trails in production. The beta is available on LangSmith Plus and Enterprise plans.
A new tool called Shepherd acts as a manager for other AI agents. It can copy workflows, undo bad decisions, and help prevent costly mistakes like deleted emails and cyber incidents.
LangChain has launched LangSmith Tool Gateway in public beta. It is a managed layer that handles login credentials, identity, and access for the tools AI agents use, so teams can control which tools each agent can call and keep a record of every action.
Google Labs unveiled Stitch Loop, a product agent that keeps improving apps after they launch. It connects a product's code and live metrics, finds problems, and creates GitHub pull requests with fixes. Waitlist sign-ups opened worldwide on the same day.
OpenAI published a second internal report on how much its researchers rely on AI coding agents such as Codex. The company says agent use has grown fast, and that the theoretical value of the tokens used now exceeds researchers' base salaries. The report also notes that the agents still often produce low-quality work on complex tasks.
ElevenLabs released two new speech-to-text models, Scribe v3 for recorded audio and Scribe v3 Live for real-time speech. The company says both make fewer errors than earlier versions, and Scribe v3 Live is also available in its ElevenAgents platform.