Agentic AI

This research introduces READY, a framework that shifts focus from benchmark performance to real-world enterprise suitability. It evaluates agents based on their ability to meet specific reliability thresholds under defined human oversight policies and acceptable costs. The method preserves individual workflow success criteria while applying a standardized qualification procedure to the human-AI system.

• Benchmark success does not guarantee deployment readiness in production environments.
• READY measures reliability and cost jointly under specific human oversight policies.
• Qualification is workflow-specific, preserving unique definitions of successful execution.
• Provides a standardized procedure to qualify agents for enterprise workflows.

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 and supports over 25,000 automated code reviews weekly. The architecture leverages isolated Firecracker microVMs, an MCP gateway, and reusable playbooks to manage workflows with scoped access and centralized auditing.

• Migrating agents from laptops to cloud reduces local resource contention and standardizes execution environments.
• Flux achieves high throughput with 130,000 tasks monthly, proving scalability for internal dev tools.
• Isolated Firecracker microVMs provide security boundaries for untrusted or semi-trusted agent code.
• Centralized auditing and scoped access mitigate risks associated with autonomous agent actions.
• MCP gateway integration allows agents to interact with multiple tools via a unified interface.

Amazon Connect Customer has made the Agentic CX Designer generally available, providing a no-code visual canvas to build AI-powered self-service experiences. This tool merges deterministic logic with large language models, allowing business users to define flowcharts, guardrails, and system integrations while the AI handles natural conversation nuances. The platform aims to reduce deployment timelines from months to weeks by enabling rapid testing and production launch of voice and digital channels.

• No-code canvas lets business users design complex agentic workflows without developer intervention.
• Combines deterministic flowcharts with LLMs for controlled yet natural customer interactions.
• Accelerates time-to-market for voice and digital self-service from months to weeks.
• Enables direct integration with backend business systems for real-time data access.

This research addresses the challenge of monitoring web agents when internal model signals like token logits are inaccessible. The authors propose prefix-level risk prediction using observable trajectory data to determine if an agent's execution path is trending toward failure. They introduce two feature representations: Macro features for cross-step behavior and Micro features for measuring intent-action consistency through repeated queries. The method relies on labeling the first uncorrected critical error rather than the final outcome.

• Enables agent monitoring without access to internal model probabilities or logits.
• Uses Macro features to summarize agent-environment feedback across steps.
• Employs Micro features to check consistency of action and state changes.
• Labels the first uncorrected critical error for more precise risk detection.
• Allows real-time intervention before minor issues become total failures.

Researchers introduce CHIME, a hierarchical memory system that addresses the credit assignment problem in self-evolving AI agents. Existing methods often confuse plan quality with execution errors by relying solely on final outcomes as feedback. CHIME improves planning accuracy at inference time by attributing credit more accurately to specific planning steps without requiring parameter updates.

• Solves credit assignment bias in self-evolving memory banks for long-horizon tasks.
• Improves planning capability at inference time without retraining or expensive search.
• Decouples plan quality from execution errors and environmental noise.
• Uses hierarchical structure to accumulate reusable experience more accurately.