Agentic AI

This paper addresses the latency bottleneck of external memory stores in language agents by embedding retrieval directly within the processing loop. While prior work limits memory access to once per turn to avoid network delays, this approach utilizes in-process stores that respond in approximately 100 microseconds. By keeping memory in-process, the system enables read and write operations on every reasoning step without the exponential latency penalties associated with networked databases.

• In-process stores reduce memory latency to ~100us, eliminating network overhead.
• Agents can perform multi-step memory reads/writes per turn without 83x latency spikes.
• Shifts memory architecture from external querying to internal, in-loop integration.
• Enables complex reasoning loops that require frequent, low-latency context updates.

This paper aggregates findings from 27 benchmark and audit papers published between 2023 and 2026 to create a unified taxonomy of LLM agent limitations. It reveals that reported benchmark improvements often mask recurring failure modes in tool invocation, multi-step planning, and long-horizon reasoning. The analysis integrates evidence across tool use, multi-agent coordination, and safety to highlight structural weaknesses in current agent architectures.

• Benchmark gains frequently obscure persistent failure modes in tool use and planning.
• First unified taxonomy covering tool invocation, planning, and long-horizon reasoning.
• Analyzes 27 papers across 19 distinct benchmarks to identify six failure clusters.
• Highlights gaps in measurement validity and multi-agent coordination reliability.
• Safety and audit findings are integrated into a single cross-cutting limitation framework.

This paper introduces AgenticAI-Supervisor, a platform designed to replace static evaluation with scalable, simulation-based reinforcement learning for LLM agents. The system decouples environment creation from execution via an API and UI, using verifiable outcomes to generate high-fidelity traces and multi-dimensional rewards. It specifically addresses reward hacking through rigorous internal state validation, as demonstrated in a customer support agent case study.

• Moves beyond static benchmarks to simulate multi-step decision-making for LLM agents.
• Decouples environment design from execution for better scalability and flexibility.
• Mitigates reward hacking by validating internal states alongside final outcomes.
• Generates high-fidelity traces to support detailed model optimization and debugging.
• Validated via a customer support agent case study showing closed-loop feedback.

Researchers demonstrate that steering vectors extracted from heading-anchor positions can exert bidirectional causal control over tool invocation in tool-augmented LLMs. This method suppresses unnecessary tool use, particularly in domains where parametric reasoning is sufficient, across five open-source models and three domains. The findings suggest stable internal representations for tool-use decisions exist, despite tools being context-only at inference time.

• Steering vectors at heading-anchors provide causal control over tool invocation.
• Effectively suppresses unnecessary tool calls where parametric reasoning suffices.
• Validated across five open-source models and three distinct domains.
• Confirms stable internal representations for tool-use exist in context-only settings.

NapMem shifts long-term user memory from passive retrieval to an active, structured action space. It organizes history into a linked pyramid of raw conversations, typed records, topic tracks, and profiles connected by provenance. The agent is trained to select and inspect specific memory granularities based on the query and intermediate evidence.

• Moves beyond passive retrieval to active memory navigation via tools.
• Uses a multi-granularity pyramid linking raw data to user profiles.
• Agent learns to inspect evidence levels before responding.
• Provenance relations ensure traceability across memory layers.

Researchers identify two inefficiencies in vanilla on-policy distillation for long-horizon agents: wasted compute on noisy tail turns and insufficient training for deep decision turns due to trajectory-level KL objectives. TurnOPD addresses this by introducing a turn-level budgeting strategy that allocates resources more effectively across the agent's interaction history. This approach aims to improve training efficiency and reduce wall-clock time while maintaining policy alignment with stronger teacher models.

• Vanilla OPD wastes resources on tail turns that provide weak KL supervision.
• Trajectory-level objectives under-train deep decision turns in long sequences.
• TurnOPD uses turn-level budgeting to optimize distillation efficiency.
• Reduces wall-clock time without sacrificing policy alignment quality.