AI & Automation Briefing - October 22, 2026
Walmart Opens Its Platform to AI Shopping Agents While Amazon Locks Them Out
Walmart launched an Agent Gateway, a Model Context Protocol (MCP) server that lets approved AI agents search inventory and complete purchases on behalf of customers. MCP was originally developed by Anthropic to connect AI agents to external data, tools, and workflows. Walmart's implementation keeps customer purchase history and personal data away from the agent itself, addressing a key privacy concern in agentic commerce. A new Agent Profiles feature, rolling out next month, lets shoppers authorize individual agents through a dedicated account profile, view agent transactions separately from their own, and set per-agent spending limits or require approval on each transaction. Walmart already has integrations with ChatGPT, Gemini, and Meta's Muse, and the Agent Gateway extends to Sam's Club as well. The contrast with Amazon is direct: Amazon has moved to block AI agents from completing purchases on its platform, a position shared by several travel and booking sites. The split signals a meaningful divergence in how major platforms are choosing to handle the growing use of AI agents by consumers, with real implications for anyone building agent-based workflows or advising businesses on how the agent economy will affect their operations.
DeLoop Trains LLM Agents to Break Out of Repetitive Tool-Call Loops Without External Rewards
Researchers have released DeLoop, a method that addresses one of the more costly failure modes in agentic workflows: agents repeatedly issuing the same tool call and receiving the same result, stalling progress and burning tokens. DeLoop identifies these loops inside an agent's own execution history and converts each one into a training pair. The repeated call becomes the rejected action, and the first action that breaks the loop becomes the preferred one. The model is then fine-tuned on these pairs using Direct Preference Optimization, requiring no task reward, labeled answer, or external judge. The problem worsens significantly when conversation history is compressed. On the CodeGym benchmark, Qwen3-8B looped in 29% of runs with full history and in 57.3% of runs under summarization, even when summaries explicitly listed prior actions. After DeLoop training, the loop rate on that model dropped from 57.3% to 6.1%, steps per task fell from 15.0 to 4.7, and token usage dropped 47% with no loss in task success. Larger models also improved on task success: Qwen3-14B gained 7.2 points and Qwen3-32B gained 7.1 points. A cross-domain variant called DeLoop-Mix pushed Qwen3-8B task success from 25.1 to 30.3. The method held up across seven memory policies and five additional benchmarks including ALFWorld and BrowseComp-Plus, where loop rates fell from 22.4% to 1.2%.
Cache Engineering Gives LLM Agents Direct Control Over KV Cache Without Recomputing Context
A new paper from Tianyu Fu and colleagues introduces cache engineering, an approach that lets LLM agents manage their key-value cache directly during long multi-step tasks. Unlike context engineering methods that trim or summarize conversation history, cache engineering lets agents drop selected cache entries from future attention and reposition retained keys without altering the trajectory or triggering expensive re-prefill operations. The researchers identify history dependence as a core serving challenge: the same token sequence can produce different cache states depending on the sequence of cache operations, making standard prefix matching unreliable for cache reuse. To address this, they built the AKV serving engine, designed to handle these stateful cache operations efficiently and reduce both storage and I/O costs for long-running agents. For teams running multi-step agentic workflows, this approach offers a path to lower latency and cost without sacrificing task accuracy.