Agentic AI

NVIDIA's Nemotron 3.5 Lightning model is now available on Amazon SageMaker JumpStart, targeting persistent agent workloads and enterprise automation. The model uses a hybrid Mixture-of-Experts architecture with 30B total parameters but only 3B active per forward pass, enabling up to 4x throughput compared to similar models. It supports up to 1M token context via DFlash speculative decoding and is distilled from the larger Nemotron 3 Ultra.

• Hybrid MoE architecture with 3B active parameters delivers ~410 tokens/sec throughput
• 30% faster task completion for domains like financial processing and cybersecurity
• Supports 1M token context window using DFlash speculative decoding
• Optimized for persistent agents and high-throughput enterprise automation
• Distilled from Nemotron 3 Ultra for efficient performance

This paper introduces Controlled Memory Interference (CMI), a framework designed to study how AI agent memories evolve through relationships like reinforcement, revision, or interference. Unlike existing systems focused primarily on construction and retrieval, CMI allows for diagnostic analysis of how new experiences impact existing memory states. The authors note that while benign accumulation has limited effects, the framework provides a structured way to examine these complex memory dynamics.

• CMI offers a diagnostic framework to analyze memory evolution in continual LLM agents.
• Moves beyond simple retrieval to study interference, revision, and reinforcement.
• Addresses gaps where multiple memories remain relevant but differ in validity.
• Provides controlled data generation for studying memory state changes.

This paper introduces a researcher agent for knowledge-graph question answering that autonomously iterates on its own prompts, rules, and tool-orchestration code. After each inference round on a validation set, the agent proposes and tests changes to improve performance. The system was instantiated on DBpedia, evolving through nine versions driven by a low-cost reasoning model before deploying the best configuration with stronger backbones.

• Agents can autonomously refine their own orchestration code and prompts without human intervention.
• Self-improvement loops significantly boost text-to-SPARQL accuracy over static tool-using agents.
• Low-cost reasoning models are sufficient to drive iterative evolution before scaling to stronger backbones.
• Validating changes against a held-out set is critical for safe agent self-modification.
• The approach demonstrates practical applicability on large-scale ontologies like DBpedia.

This survey examines co-evolution in agentic systems, where multiple agents and their environment mutually adapt through dynamic pressure. It proposes a three-stage taxonomy tracking how systems shed static, human-engineered constraints in favor of continuous self-improvement. The framework covers agent-agent interactions like adversarial or collaborative adaptation, and agent-environment loops involving adaptive tasks and feedback.

• Co-evolution overcomes static learning limits of single-agent self-improvement by using multi-component adaptive pressure.
• Taxonomy traces system progress from human constraints to fully self-directed evolutionary loops.
• Agent-agent co-evolution includes adversarial, collaborative, and organizational adaptation mechanisms.
• Agent-environment co-evolution integrates adaptive tasks, feedback, and dynamic interaction contexts.

This study benchmarks nine LLMs against a Perfect Bayesian Equilibrium in 9,840 autonomous supply chain negotiations involving private demand information. While agents reached agreements in 98.9% of cases and captured 95.4% of the theoretical surplus, they averaged 2.98 rounds compared to the benchmark's 1.25. This negotiation delay eroded 21-34% of the total surplus, indicating that current capabilities significantly impact efficiency despite high success rates.

• LLMs achieve 98.9% agreement rates, proving reliability in autonomous procurement tasks.
• Negotiation delay is the primary efficiency killer, eroding 21-34% of potential surplus.
• Current LLMs require roughly double the rounds of optimal game-theoretic strategies.
• Capability tiers exist among models, directly correlating to surplus preservation.
• Delegating negotiations introduces hidden costs through extended bargaining rounds.

This paper addresses the failure of fragmented AI pilots in hospitals by proposing a compliance-first orchestration framework. It introduces a catalogue of agentic AI patterns designed to stabilize mission-critical workflows like triage and revenue cycle management. The goal is to prevent operational fragility and technical debt as healthcare institutions scale AI adoption toward production environments.

• Shift focus from isolated chatbot pilots to governed, multi-agent ecosystems.
• Use the provided pattern catalogue to standardize agentic workflows in hospitals.
• Prioritize compliance and risk governance to avoid scaling architectural missteps.
• Target mission-critical areas like triage, documentation, and revenue cycles.
• Framework aims to reduce technical debt and operational fragility in production.

A new audit reveals that embedding-cosine similarity thresholds often misalign with semantic intent in agent systems. While these gates measure wording changes, they fail to detect meaning-breaking mutations like instruction reversals, which can occur with minimal lexical edits. The study found that a production drift guard caught zero critical safety violations while incorrectly approving one item.

• Cosine similarity measures lexical overlap, not semantic equivalence.
• Single-word reversals can bypass high-similarity safety gates.
• Current drift guards may miss critical meaning-breaking mutations.
• Relying solely on cosine thresholds creates false sense of security.
• Audit semantic gates against specific mutation types, not just drift.