Production agent systems: models, infrastructure & deployment

Rapid release cycle continues with Meta Muse Spark 1.1, Qwen3.7-Max, Claude Opus 4.8, NVIDIA Nemotron 3 Ultra, Google Agent Executor, Agentix, Microsoft MAI. Qwen-AgentWorld Language World Models (397B MoE) outperforms GPT-5.4/Claude 4.8, open-source. Xiaomi HarnessX +44% gain. Microsoft self-healing cloud agents. Sail Research $80M for long-horizon agent infrastructure. Ornith-1.0 learns own RL scaffold. KAIST study reveals AI agents consume up to 136.5x energy per query vs standard LLM, with GPUs idle 54.5% of time — critical efficiency bottleneck. Practical tool selection guide shows RAG-MCP integration yields 13.62% to 43.13% accuracy gains. First real-world agentic ransomware (JadePuffer) validates security threats. LLM-as-a-Verifier: SOTA 78.2% SWE-Bench, 86.5% Terminal-Bench, training-free. Enterprise AI agent production guide: 95% failure rate (MIT NANDA), five pillars. CompactionRL — joint optimization of task execution and summary generation for long-horizon agents; gains on SWE-bench and Terminal-Bench; scaled to GLM-5.2. AgenticAI-Supervisor — simulation environment for agentic RL, decoupled env creation from execution, deterministic reward shaping, reward hacking mitigation. GIRA — scalable MARL with population-adaptive training and interaction-aware role generation; zero-shot generalization to unseen team sizes. Gartner 40% adoption vs 40% cancellation framing — reframes agent deployment as infrastructure/governance issue. Mistral Robostral — model family for planning and tool use with explicit planning traces, hierarchical attention, adversarial training, and error recovery; open-weight. TurnOPD — turn-level on-policy distillation for long-horizon agent training. Prime Intellect raises $130M for open agent infrastructure. New: Multi-agent robotic control with onboard VLMs (3B params) using Megamind orchestration, fine-tuning improved inspection accuracy 76.7%→91.5%, open-source sim environment. New: Automating Design of Embodied Agent Architectures — AgentCanvas runtime, KDLoop search. New: Escalation Framework for LLM Adaptation — formalizes three handles (input, parameters, harness). New: Meta Muse Spark 1.1 — context compaction for long-horizon multi-agent workflows. New: UniClawBench — capability-driven benchmark for proactive agents. New: StreamForce — AI agent platform built on a live model of production environments. New: IBM Bob multi-agent update with subagent isolation and Bobalytics cost analytics — 40% cost reduction claim. New: LLMOS — OS for small local model (Ornith 35B MoE) that autonomously handles software environment setup. New: Agentic Search in the Wild — large-scale log analysis of 14.44M agentic search requests. New: Making agentic token costs visible in production — Datadog piece on tool catalog pruning, session history capping, retrieval deduplication. New: LangChain OpenWiki Brains — proactive agent memory connecting to Gmail, Notion, git, Twitter. New: LM4Plan@ICML26 workshop signals — Sturtevant claims 99%+ planning success via training, Kambhampati verifier perspective, Brown compute-aware evaluation curves. New: CellPilot — agentic workflow (plan, execute, inspect, revise) boosts small SLM (8B) from 0.39 to 0.89 on single-cell annotation, closing gap to GPT-4o. New: Harness Engineering Is the Operating System for AI Software Delivery — reframes agent deployment as harness engineering (context, skills, boundaries, verification, learning), not model capability; aligns with production failure rate data. New: Enabling Black-box RPC-API Testing with Multi-Agent RL+LLM at Alibaba Cloud — hierarchical MARL (group, operation, parameter, dependency agents) + fine-tuned LLM for constraint-aware input generation; 77.67% operation coverage, 57 bugs found. New: Learning group collaboration for efficient multi-agent — new MARL paper on group collaboration for efficient coordination. New: Feasibility-Shielded Agentic AI Framework for 6G Self-Healing — combines digital twin, LLM agents, and deterministic validation layer (schema + lookahead simulation) achieving 100% action success rate; small open-source models (down to 1GB) match proprietary ones when constrained; practical pattern for safe agent deployment in production networks. New: Practical local agent tracing guide (Strands, Ollama, OpenTelemetry) — model choice dominates tool-calling reliability, two-pass structured output for small models, domain-specific trace attributes. New: Prime Intellect Verifiers v1 — composable tasksets/harnesses/runtimes for agentic RL training, linear trace growth, interception server for reward hacking mitigation, direct prime-rl training support, Harbor compatibility. New: TRACE — capability-targeted agentic training system from Stanford; contrastive capability analysis + targeted synthetic envs + MoE composition; 27B model beats GPT-5.2-Codex on SWE-bench. New: Cynative — open-source security research agent with read-only enforcement by default, runtime-resolved action classification from cloud provider APIs, verifier refutes findings without writing. New: ATHENA-R1 — RL-trained agent for medical treatment reasoning; two-stage (synthetic data + RL); tool-use skill matters more than model scale; 94.7% accuracy on FDA questions, 82.9% on patient-specific decisions. New: Anti-patterns for building AI agents — practical guidance on premature multi-agent, tool sprawl, no memory design; aligns with production failure data. New: PalmClaw — native on-device agent framework for mobile phones, exposes device capabilities as tools with explicit arguments, 11.5% task success improvement, 94.9% faster completion, execution boundary safety design. New: Sakana Fugu adds Nemotron — learned routing architecture (CMA-ES coordinator + RL-trained Conductor) for multi-model orchestration; challenges single-model supremacy, offers practical path for enterprises using open-weight specialists. New: SearchOS-V1 — explicit state management (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) and pipeline-parallel scheduling for multi-agent information-seeking; strong results on WideSearch and GISA; addresses agent loop problem. New: LLMs and Agentic AI Systems for Smart Grids — tutorial with solver-grounded design, verification gate, 7.5-9.5x unmet energy reduction, 52.3% violation reduction; practical evaluation framework.