AI & Automation Briefing - October 21, 2026
More Memory Can Make LLM Agents Perform Worse, New Research Shows
A paper from researchers Wang, Panaganti, and Zhou identifies a failure mode in LLM agent memory design called error reproduction. When agents store not just raw trajectories but also the rules and abstractions derived from them, faulty abstractions get retrieved, shape future executions, and generate new abstractions that repeat the original errors. The researchers propose a design called Anti-Persistence, which retains trajectories and outcomes but reconstructs task-specific guidance only when needed and discards it after use. Across multiple benchmarks and model backbones, Anti-Persistence achieved higher success rates on held-out tasks and the lowest average LLM-call cost compared to other memory designs tested. The core finding is that accumulating experience and permanently storing interpretations of that experience are two separate design decisions, and conflating them introduces compounding error risk.
Figma Expands MCP Support While Its Co-Founder Flags Protocol Limitations
Figma announced 20 new MCP partners at its first Config India conference, including ClickUp, Linear, Notion, Mistral, and JetBrains' Junie, even as co-founder Dylan Field continues to voice reservations about the protocol itself. MCP is an open standard that defines how AI agents connect to external data sources and tools. Field has publicly questioned why Anthropic, which created MCP, does not expose its own artifacts and cloud resources through the standard, framing the issue as one of asymmetric context sharing rather than simply open versus closed. At a media briefing, he stated that MCP has significant room for improvement and is currently structured in ways that favor frontier labs. He also explained that Figma limits how much data it exposes through MCP because broader exposure increases the risk of abuse, even when most users have no malicious intent.
HybridResearcher Trains a 4B Agent to Reason Across Structured and Unstructured Data Sources
Researchers have introduced HybridResearcher, a 4-billion-parameter AI agent designed to answer questions that require pulling evidence from both structured sources, such as databases, and unstructured sources, such as web text. Current frontier agents perform poorly on this type of cross-source reasoning, a gap exposed by benchmarks like HybridDeepResearch. The training approach starts with validated single-source tasks and constructs paired cross-source examples by building the missing source around a shared bridge entity from the original task output. Samples are only kept when strict validation confirms the synthesis is sound. For operations teams building research or data retrieval workflows, this points toward agents that can handle the mixed-data-source complexity common in enterprise environments.