Agentic AI

DoorDash engineered Ask DoorDash by integrating large language models with specialized agents and MCP-based tooling, supported by an intelligence layer handling persistent consumer memory and live backend data. This hybrid approach moves beyond pure generative text to execute precise shopping tasks with real-time context. Early metrics indicate a 24% lift in checkout conversion and a 17% increase in basket size, driven by improved intent accuracy through memory-backed sessions.

• Combines LLMs with specialized agents and MCP tooling for precise task execution.
• Persistent consumer memory and live backend data drive intent accuracy.
• Resulted in 24% higher checkout conversion and 17% larger average baskets.
• Architecture prioritizes deterministic tool use over pure generative responses.

The AWS Lambda console now offers a one-click prompt to configure coding agents with AWS Serverless skills and the Serverless Model Context Protocol (MCP) server. This update embeds serverless best practices directly into the agent configuration, removing the need to manually navigate documentation for setup. Developers can initiate this streamlined process at any stage of their Lambda journey, from initial exploration to creating their first function.

• Eliminates manual configuration friction for coding agents in the Lambda console.
• Automatically installs AWS Serverless skills and Serverless MCP server.
• Embeds serverless best practices directly into agent workflows from day one.
• Available across all Lambda console entry points for new and existing users.

Public battery aging datasets often suffer from inconsistent formats and scattered metadata, making them difficult to use for health management. BatteryLake addresses this by implementing a governed data lakehouse that uses LLM agents to extract metadata and synthesize dataset-specific converters. This agentic, physics-grounded framework ensures outputs are tied to verbatim evidence, automatically abstaining when no supporting data exists.

• Solves fragmentation in public battery aging data through automated, reproducible curation.
• Uses LLM agents to generate converters grounded in verbatim evidence from source papers.
• Mitigates hallucination risks by having agents abstain when no metadata evidence is found.
• Provides benchmark-ready assets by handling domain-specific semantics of electrochemical time-series.

Multi-agent and memory-augmented LLM systems often place coordination content, shared state, prior discussion, tool outputs, summaries, and role instructions, inside the same finite prompt used for the current task. This creates a practical allocation problem: every token spent on coordination is unavailable to task instructions or evidence when a call is assembled under a fixed context budget. The authors introduce the Roundtable Context Window Test (RCWT) to measure this displacement by varying coordination content while controlling total budget, position order, task family, and scoring.

• Coordination tokens directly displace task-relevant evidence in fixed-context calls.
• RCWT isolates this effect by varying coordination content while holding budget constant.
• Protocol controls for position order, task family, and scoring to ensure valid metrics.
• Current commercial models show baseline recall retention under moderate displacement at W=4096.

This research addresses the latency and data-sensitivity constraints of deploying large language models for industrial closed-loop control by fine-tuning compact Small Language Models like Qwen2.5-3B. The proposed framework replaces opaque, high-latency cloud inference with a local agent that generates control policies from natural language specifications. To ensure safety, the system employs a multi-agent self-correction loop where a plant-aware validator, such as a digital twin, checks candidate actions before execution.

• Fine-tuned SLMs offer a viable path for low-latency, edge-native industrial control automation.
• Natural language policy generation reduces manual redesign effort for control systems.
• Validator-guided correction loops mitigate hallucination risks in autonomous decision-making.
• Digital twin integration enables safe pre-execution verification of AI-generated actions.