Agentic AI Moves to Production
As agentic AI systems are deployed in production environments, a new “harness engineering” playbook is emerging that combines LLMs with tool APIs, persistent memory, and policy layers. Best practices for deploying agents in logistics and inventory management now focus on operational safety, as even small changes in prompts or tools can create unpredictable behaviors. Enterprises are also being advised to demand open standards to avoid vendor lock-in as AI agents evolve into full platforms.
- To ensure interoperability and prevent vendor lock-in, several open standards are emerging for agentic AI. These include the Open Agentic Schema Framework (OASF) for defining agent capabilities, the Agent Connect Protocol (ACP) for communication, and the Model Context Protocol (MCP) for standardizing connections between AI models and tools. Another key standard is the Open Standard for AI Agents (OSSA), which provides a contract layer to define an agent's identity, trust boundaries, and governance in a vendor-neutral YAML specification. - In warehouse automation, AI agents are moving beyond simple task execution to orchestrate entire workflows, from inventory tracking to order fulfillment. For example, Walmart uses agentic AI with computer vision and sensors to monitor stock levels and trigger restocking orders automatically. These systems can optimize warehouse layouts, direct autonomous mobile robots in real-time, and even reduce energy consumption by managing lighting and climate control. - The concept of "harness engineering" is critical for deploying agentic AI safely and effectively. This involves building a system around the AI model that includes prompt design, tool selection, and execution flow to optimize for performance, cost, and latency. Key techniques in harness engineering include self-verification, where agents test their own work, and providing agents with clear context about their operating environment to reduce errors. - For safe and reliable deployment in production, agentic AI systems require durable execution frameworks to manage long-running processes and recover from failures. Running agents in containerized sandboxes with strict resource limits and network access helps prevent unsafe code execution. Treating every agent action with the same scrutiny as a financial transaction, including requiring permissions and allowing for human escalation, is a recommended practice to manage operational risk. - The shift to agentic AI is changing the role of engineers from hands-on coding to a model of "delegate, review, and own." In this model, AI agents handle the initial execution of tasks like implementation and testing, while human engineers review the output for correctness and alignment with system architecture, ultimately retaining ownership of the outcomes. - In manufacturing, agentic AI coordinates multiple agents across production, maintenance, and logistics to make real-time decisions. For instance, if a sensing agent detects a potential equipment failure, a maintenance agent can automatically create a work order, check for spare parts, and schedule a repair to minimize disruption, while simultaneously updating delivery timelines in the supply chain system. - A significant challenge in deploying agentic AI in supply chains is the probabilistic nature of these systems, which contrasts with the deterministic behavior of traditional ERP systems. This variability requires robust data validation and clear governance to define which decisions an agent can make autonomously and when human intervention is necessary. - According to a 2025 global survey of nearly 1,500 IT leaders, 96% of organizations plan to expand their use of AI agents, with 84% believing they are essential for staying competitive. However, significant barriers to adoption remain, including data privacy and compliance (cited by 53%), integration with legacy systems (40%), and a shortage of AI-skilled talent.
