Agentic AI

Grab has deployed LLM-Kit, an internal framework that standardizes and manages over 500 AI agent services. The system centralizes infrastructure, tool discovery, and secret handling, drastically reducing deployment time from two weeks to one hour. This approach allows for flexible model integration while maintaining strict operational control across the fleet.

• Standardizes 500+ internal agent services to reduce fragmentation.
• Cuts deployment cycle from two weeks to one hour via automation.
• Centralizes secret handling and evaluation for consistent security.
• Enables runtime tool discovery without sacrificing operational control.

AWS has launched a new interface that allows users to interact with AI agents using an email-like paradigm. The tool, exemplified by the 'Pizza Bot' demo, enables local execution of agents while providing a familiar inbox-style experience for receiving reports and posing questions. This approach aims to streamline communication between human operators and autonomous AI systems.

• AWS now supports an email-style UI for managing AI agent interactions.
• Agents can run locally while maintaining a centralized inbox interface.
• Simplifies reporting and querying for AI workflows with familiar UX.
• No direct database infrastructure changes required for this feature.
• Focuses on human-agent communication rather than backend scaling.

IBM Research introduces a method to verify if large language model agents can reliably repeat successful tasks rather than relying on luck. The approach leverages adaptive learning techniques to stabilize agent behavior across multiple runs. This addresses the common issue where agents perform well in isolation but fail to maintain consistency in production workflows.

• Agents often succeed by chance; this work focuses on repeatable reliability.
• Uses adaptive learning to stabilize decision-making paths in LLM agents.
• Critical for production systems requiring deterministic outcomes.
• Reduces evaluation noise when testing agent performance over time.

Alex Porcelli addresses the accountability risks of non-deterministic LLM outputs in high-stakes enterprise scenarios. The proposed architecture integrates DMN decision models with agent skills and NeMo guardrails to enforce deterministic behavior. This approach allows business leaders to control decision logic while engineers manage the underlying architectural governance.

• Combine DMN models with LLMs to ensure auditable, deterministic outcomes in agentic workflows.
• Use NeMo guardrails to enforce safety and compliance boundaries around agent actions.
• Separate decision logic ownership (business) from architectural governance (engineering).
• Mitigate non-deterministic output risks in high-stakes enterprise AI deployments.