Agentic AI

DoorDash has migrated its engineering agent workloads from local developer machines to its internal Flux cloud platform. This move enabled the automation of 130,000 engineering tasks in a single month, including over 25,000 weekly automated code reviews. The system leverages isolated Firecracker microVMs, an MCP gateway, and reusable playbooks to execute these workflows with scoped access and centralized auditing.

• Migrating agent workloads from laptops to cloud infrastructure significantly scales automation capacity.
• Flux platform processed 130,000 engineering tasks monthly, demonstrating high-volume operational efficiency.
• Security is enforced via Firecracker microVMs, providing isolated execution environments for agents.
• Centralized auditing and scoped access control are built into the agent workflow invocation layer.
• The architecture supports multiple invocation surfaces and reusable playbooks for flexible task execution.

AWS has released the Agent Registry as generally available, offering a private, governed catalog for agents, tools, skills, and MCP servers. It enables teams to discover existing AI capabilities rather than rebuilding them, with access via the console, CLI, SDK, and as an MCP server for IDE integration. The service supports semantic and keyword search to help builders find resources within Amazon Bedrock AgentCore, Quick, and Kiro IDE.

• Centralizes discovery of agents, tools, and MCP servers in a private, governed catalog.
• Reduces redundancy by allowing teams to reuse existing AI capabilities.
• Integrates directly into IDEs via MCP server for seamless builder workflows.
• Available through console, CLI, SDK, and within Bedrock AgentCore and Kiro IDE.

Amazon Quick now allows users to search, browse, and enable agents and MCP servers from the AWS Agent Registry directly within the interface. Connection details are automatically populated from the registry, removing the need for manual configuration. This integration enables technical teams building with Amazon Bedrock AgentCore to share resources with business users across chat, apps, and deep research workflows.

• Search and enable AWS Agent Registry resources directly in Amazon Quick without manual setup.
• Connection details auto-populate, reducing configuration friction for end users.
• Bridges gap between Bedrock AgentCore builders and Quick business users.
• Shared agents work across chat, apps, flows, and deep research contexts.

Current personalized agents rely on text transcripts and captions, which fail to capture perceptual data like voice tone or facial expressions. The authors demonstrate that caption-based retrieval recovers only 0.11 recall compared to dedicated encoders for non-nameable signals. Their approach decomposes memory into a vision-language model for grounding and a dedicated encoder for perceptual storage.

• Text-only memory discards critical perceptual signals like voice and face recognition.
• Caption-based retrieval performs near chance on non-nameable user attributes.
• Proposed method uses VLM for context grounding and dedicated encoders for memory.
• New architecture enables persistent multimodal user profiles beyond simple text.

Terminal-Bench-LILT introduces a multilingual coding benchmark with 300 tasks across ten languages, addressing the English-centric bias in current AI evaluations. The tasks focus on region-specific challenges like internationalization, encoding, and cultural conventions, authored by native speakers. Tests on six leading models show a maximum pass rate of only 63.1%, highlighting significant performance drops in non-English contexts.

• Current coding agents fail significantly on non-English tasks, maxing out at 63.1% pass rate.
• Benchmark covers 10 languages including Arabic, Japanese, and Hindi with native-authored tasks.
• Focuses on real-world issues like text normalization and cultural conventions absent in English.
• Reveals that top-tier models are not yet ready for global, multilingual deployment.

Amazon Redshift now supports the Agent Toolkit for AWS, allowing AI coding assistants like Claude Code and Cursor to manage data warehouses via the AWS MCP server. This integration provides authenticated API execution and specific Redshift skills to help agents generate SQL, explore schemas, and optimize materialized views. The feature aims to reduce query errors and streamline data loading and migration tasks through curated reference materials.

• Enable AI agents to execute authenticated AWS API calls for Redshift management tasks.
• Leverage Redshift skills for accurate SQL syntax, schema discovery, and best practices.
• Supports popular AI tools including Claude Code, Kiro, and Cursor.
• Covers data loading patterns, materialized views, and extensions like Pivot.