Agentic AI

AWS Database Migration Service Schema Conversion now supports AI agent automation through the AWS MCP Server, enabling integration with coding agents like Kiro, Claude Code, and Cursor. These agents can execute complete migration workflows using natural language, autonomously creating projects, browsing metadata, converting schemas, and generating assessment reports. A dedicated skill, dms-schema-conversion, provides agents with predefined procedures and operational rules to reduce trial-and-error loops during migration tasks.

• Integrate DMS Schema Conversion with IDE-based AI agents via the AWS MCP Server.
• Agents can autonomously run full migration workflows using natural language instructions.
• The dms-schema-conversion skill enforces best practices and reduces trial-and-error.
• Supports automated schema conversion, metadata browsing, and report generation.
• Extends automation to code objects like stored procedures with agent assistance.

This research paper introduces the Context Graph, a live relational data structure that models enterprise entities and their state transitions over time. It proposes a Delta Detection Engine to monitor changes and a Proactivity Scorer to rank insights by urgency and relevance. An LLM-powered Surfacing Layer then delivers these proactive recommendations to workers before they explicitly ask for them.

• Moves enterprise AI from reactive query-response to proactive insight delivery.
• Uses a Context Graph to track entity relationships and state changes over time.
• Delta Detection Engine continuously monitors for meaningful state shifts.
• Proactivity Scorer ranks insights by urgency, relevance, and user persona.
• Aims to boost productivity by surfacing actionable info before human requests.

Google has released a library of Agent Skills compatible with the Stitch MCP server, designed to integrate with major coding agents like Claude Code, Cursor, and Codex. These skills adhere to the Agent Skills open standard, enabling consistent interaction across different AI development environments. The release includes a quick start guide for installing the plugin suite globally or per repository via CLI.

• Enables cross-agent compatibility for Stitch design tasks using the open Agent Skills standard.
• Supports popular coding agents including Claude Code, Cursor, Gemini CLI, and Codex.
• Provides CLI-based installation methods for global or sparse repository plugin management.

This paper addresses the transition of enterprise LLM applications from prompt-driven prototypes to production-grade systems by introducing a harness-engineering approach. It shifts deterministic behavior into code, manifests, and validation artifacts, establishing a replaceable composition boundary while keeping source-backed claims as the authority for runtime answers. The authors validate this architecture using a public dataset of five Korean corporate groups to demonstrate improved traceability and auditability.

• Shifts deterministic logic from fragile prompts to structured code and schemas for reliability.
• Establishes a replaceable composition boundary to decouple orchestration from model logic.
• Ensures runtime answers remain grounded in source data rather than model hallucination.
• Validates the pattern on real-world corporate data to prove enterprise-grade auditability.

TencentCloud’s new open-source project implements a fully local, four-tier memory architecture for AI agents, eliminating external API dependencies. It combines symbolic short-term memory to condense tool logs with layered long-term memory that structures conversations into personas and scenes rather than flat vectors. Integrating this system with OpenClaw reportedly reduces token consumption by over 61% and boosts task pass rates significantly.

• Uses symbolic short-term memory to compress tool logs into Mermaid diagrams, reducing token overhead.
• Implements layered long-term memory to structure conversations into personas instead of flat vector piles.
• Fully local execution ensures zero external API dependencies for enhanced privacy and control.
• Integration with OpenClaw yields 61.38% token reduction and 51.52% relative pass rate improvement.
• PersonaMem accuracy increases from 48% to 76% through structured memory distillation.

CausalDS addresses the gap between symbolic causal reasoning benchmarks and standard data analysis evaluations by testing LLMs within realistic agentic workflows. The benchmark utilizes systematically generated synthetic causal structures rather than limited templatized variations, ensuring greater diversity in evaluation scenarios. This approach allows for a principled assessment of how agents handle data generation and causal inference simultaneously.

• Existing benchmarks lack realistic data analysis combined with principled causal structures
• CausalDS uses systematic generation of novel synthetic causal structures for diversity
• Evaluates LLMs as integrated agents performing both reasoning and tool use
• Fills the gap between symbolic reasoning and practical data science workflows