Agents

LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning

600 articles · page 2 of 13

Agents

Adding a motion planner lifts robot household task success from 57% to 77%

Roboticists at NVIDIA Research, Georgia Tech and the University of Utah combined large vision-language model agents with task and motion planning, the classical technique that checks whether a grasp, pose or trajectory is physically feasible before a robot moves. On simulated kitchen tasks with a two-armed mobile robot, the hybrid reached 77.2% absolute success, against 57.2% for a direct tool-calling agent and 61.1% for a graph-based variant, while using fewer tokens per minute of simulated time. The authors argue that constraint checking, not just commonsense reasoning, is what these agents were missing.

·3 min
Agents

Customer-service AI agents can improve from their own mistakes after deployment

The authors present CARE, a recursive self-improvement framework that lets a customer-service AI agent learn from its own production failures and adapt to changing business policies. After a one-time cold start on historical data, the agent serves live traffic alongside the incumbent frontier model; its errors are identified through blinded independent review, attributed to the model or other components, and used to update the model via on-policy self-distillation. In simulations, candidates must match or exceed the deployed version's performance before human approval for release.

·3 min
Agents

AI models find the best team deal in only 14% of conflicts

Researchers at Stanford, Oxford and MIT built the Collaborative Intelligence Benchmark to measure how well an AI agent behaves as a member of a team rather than on its own. Each task runs three times: the model alone, the model in a normal team, and the model in a team where a scripted partner either pushes for something unsafe or needs help. Models are scored only on tasks they solve alone, so failures can be attributed to teamwork rather than ability. Across 18 models, the best scored 82.2 out of 100, and models of nearly identical general capability differed by up to 11 points as teammates.

·3 min
Agents

AI web agents learn long tasks by reusing skills salvaged from failures

Feng et al. tackle a practical obstacle in training web agents: when a browsing task takes dozens of clicks, the agent almost never stumbles onto a success by chance, and sparse rewards mean failed runs are usually thrown away. Their framework, SCOPE, treats those failures as raw material, mining them for reusable local skills (reaching an order page, applying a filter) and for mistakes worth avoiding. Across three web benchmarks the approach outperformed prior agent-training methods.

·3 min
Agents

Most AI agent tools that delete data run without asking a human

The Model Context Protocol lets AI agents call external tools, including ones that delete databases, send messages or move money, and its specification says there should be a human able to deny such calls. Yao et al. measured both sides of that promise from source code: the registry servers that offer destructive tools, and the host applications and agent SDKs that decide whether to ask the user. They find the human is present when the host prompts and almost never otherwise.

·3 min
Agents

AI agents get better at software tasks by reusing their own verified work

NorthCUA is an integrated system that generates executable desktop and web tasks, runs them in a harness that records every decision and checks outcomes, then feeds the verified trajectories back into training. The authors report that the resulting Nano, Mini and Pro models lead open-source systems on GUI grounding, desktop execution and web interaction under their reported evaluation settings. A faster variant trades some quality for lower latency in web workflows.

·3 min
Agents

AI agents picking from a menu succeed more often with less training

Chen et al. propose RLiD, a reinforcement learning setup that skips text generation during training rollouts. The model reads the observation plus a list of available actions and outputs a probability for each action label in one forward pass, so every sampled action is executable and nothing needs to be parsed. The authors pair this with an analysis of why word-by-word credit assignment is noisy, and with a distillation step that transfers the learned policy back into an ordinary text-generating model.

·3 min
Agents

Optimizing AI agents for cost as well as quality improves both

ParetoSaddler is a method for automatically rewriting the scaffolding around a language model agent, its prompts, tools, middleware and control logic, while optimizing answer quality and token cost at the same time. The authors show that preserving candidate harnesses which improve only one of the two objectives, and screening them with a statistical test over repeated runs, yields harnesses that beat quality-only optimization on quality as well as cost.

·3 min
Agents

Terminal prompts make AI agents report success without doing the work

LLM agents run command-line tools assuming each call returns with the requested action finished. Many commands, such as a database migration or a key rotation, ask for confirmation first. The authors show that how the agent's execution harness handles that question decides whether the action happens at all, and that a leaderboard's ranking of models partly reflects this plumbing rather than the model. Their system, RECEIPT, reads from the operating system what each process is waiting for and hands that state back to an unchanged agent.

·3 min
Agents

AI sales agents that learn from past chats lift conversions 16%

Researchers at South China University of Technology and Alibaba Group built SPAR, a system for B2B outbound sales agents that keeps memory across conversations and revises a shared playbook of sales strategies without retraining the underlying model. In a live A/B test with more than 200,000 merchants on the 1688 platform, the authors report a 15.8% lift in conversion and a 35.8% increase in high-intent merchants. Domain experts rated its sales quality 4.62 out of 5, the highest among the methods compared.

·3 min
Agents

Reusable parts let a single fix repair many broken AI training environments

AI agents learn by acting inside simulated software, and those simulated environments are costly to build and often quietly broken. Jiang et al. propose assembling them from reusable pieces, which they call atoms, so that when an environment fails, the faulty piece can be fixed once and every environment built from it improves. Training on the resulting corpus doubled one model's score on a standard office-automation benchmark.

·3 min
Agents

AI models share core values but differ on when to stop self-editing

A single-author study gave six commercial language models an empty system prompt and a loop: read the prompt, draft an improvement, test it on fresh conversations, then commit the change or declare the job done. Both Claude models declared convergence in every run. GPT-4o did so in 11% of runs, one of which ran 101 rounds. The values the models wrote into their prompts converged on a small shared set and stayed there through repeated edits.

·3 min
Agents

Reusing past decisions in AI agents reduces token cost by up to 88% and improves success

The authors introduce HarnessMM, a technique that reduces redundant reasoning in AI agent harnesses by reusing actions whose semantics have been fixed by prior deliberation. On HumanEval, it maintained Codex success at 90.82% while reducing token consumption by 29.54%; on DeepSeek Harness, it improved success from 71.43% to 85.71% while cutting tokens by 87.93%. An ablation shows that both reuse and callback removal contribute to savings.

·2 min
Agents

Structured backtracking beats reflection for long-horizon language agents

The authors present the Verify-Backtrack Loop (VBL), a training-free extension of ReAct that keeps task decomposition as a bounded goal stack and makes failure diagnosis a structured operation. Evaluating three language models on AppWorld, ALFWorld, and ScienceWorld, they find VBL outperforms ReAct and reflection-based control on every backbone-benchmark pair, with gains up to 22.07 percentage points on AppWorld.

·3 min
Agents

RL method helps AI agents learn high-level task strategies from sparse rewards

The authors introduce TalentRL, a framework where large language model agents learn both a high-level strategy (talent) and fine-grained skills by using matched-intervention reward contrasts to isolate the contribution of strategy selection. Across seven question-answering benchmarks, it improves success rates by up to 9.32% over supervised fine-tuning alone, and achieves state-of-the-art results on ALFWorld (99.3%) and WebShop (90.7).

·2 min
Agents

New memory architecture lets legged robots remember maze layouts while exploring

The authors propose a spatially indexed memory module for reinforcement learning that allows a legged robot to store and recall task-relevant information at specific positions in the environment. Tested in simulated and real-world mazes, the system matches exploration coverage of standard recurrent policies while uniquely preserving the maze layout, cutting reconstruction error by a factor of three.

·3 min
Agents

Coding up definite operations improves AI agent consistency

The authors show that transforming well-specified skill operations into executable programs reduces reliance on LLM reasoning and improves task performance. Their framework, SkillCompiler, uses a compilation strategy to identify compilable operations and a reuse mechanism to avoid redundant program synthesis. Across extensive experiments, it consistently outperformed state-of-the-art methods, achieving up to a 32.33% improvement.

·2 min
Agents

Single-image 3D generation creates assets that move and deform realistically

The authors introduce PhysAg3D, a framework that combines a 3D generative model with a vision-language agent to produce interactive assets from a single image. The agent infers semantic parts and physical priors, the generator creates a textured mesh, and the agent uses geometric and physical tools to construct joints, deformable regions, and surface bindings guided by a hierarchical interaction representation. Results on real photographs and benchmarks show plausible interactions, both rigid articulation and compliant deformation, while retaining image-aligned appearance.

·3 min
Agents

AI assistants that learn separate style modules from response feedback outperform fixed-prompt personalization

Shen et al. introduce Persona Learning from Feedback (PLF), which learns reusable preference-grounded assistant personas as independent LoRA adapters on a shared frozen backbone. During training, each adapter generates responses and receives feedback; a frozen self-teacher provides token-level supervision on the adapter's own prefixes. A feedback gate focuses updates on responses needing improvement. After learning, a supervised router selects or composes adapters for each request without user-specific finetuning. In experiments, PLF improves model-judged preference fit over persona prompting, persona distillation, and fixed-policy baselines in both single-turn personalization and multi-turn tutoring.

·2 min
Agents

Game engine expertise and verification feedback boost AI 3D scene completion from 18% to 48%

The authors present CineU, an agentic system that controls Unreal Engine via natural language instructions to construct, edit, and choreograph scenes. They evaluate on compound tasks and show that combining specialized engine knowledge (code patterns for common operations) with multi-stage verification and function-level repair substantially outperforms baselines without these components. The best configuration completes 47.8% of tasks, compared to 18.4% without knowledge and 16.8% without verification.

·2 min
Agents

Multi-agent AI systems improve by learning from execution traces

The authors introduce MAS-Loop, a framework that pairs an Orchestrator with a trajectory-conditioned Refiner. After each task, the Refiner inspects the execution log, localizes failures, and proposes bounded patches to agents, information flow, or execution configuration. The updated system is re-executed on the same task, and the score difference provides learning signals for both policies, enabling co-evolution. Experiments show consistent improvements across interaction, multi-hop reasoning, and general reasoning benchmarks, and the resulting dataset supports improved test-time MAS selection.

·3 min
Agents

AI agents improve on specialized tasks with 17–38% fewer training runs

The authors propose EMMA (structured Episodic Memory for Multi-Agent workflow synthesis), which treats the entire configuration of a multi-agent system (controller routing, agent prompts, procedural skills) as episodic memory. EMMA introduces explicit scope assignment and backbone-independent guards for each update, leading to consistent improvements over strong baselines across five backbones and three benchmarks, with up to 32.5% performance gain and 17–38% fewer training executions.

·3 min
Agents

AI agents halve simulated harmful outcomes when safeguards weigh consequences

The authors propose CASEL, a runtime layer that maps proposed tool calls to one of six intervention modes, balancing failure probability, consequence, reversibility, ambiguity, and evidence staleness. In a 20-seed simulation with 120,000 test actions per seed, CASEL roughly halved harmful-side-effect trajectories relative to an uncertainty-ranked gate, at equal compute and with the same intervention menu. A simpler risk-ranked gate achieved slightly lower harm but required 45 percent more human approvals.

·2 min
Agents

Coding agents working in teams can outperform expert-written AI chip code

Cao et al. introduce Evolving Agent Teams, a multi-agent framework in which a central supervisor monitors progress, diagnoses why a team is stuck, and restructures the team's roles, all grounded in a persistent file system. On four CUDA kernel optimization tasks, the framework outperforms strong single-agent and static-team baselines under matched token budgets, and with extra budget surpasses the production FlashInfer MLA decode kernel by 1.21x mean speedup.

·2 min
Agents

AI agents can learn from imagined outcomes without generating images

The authors introduce Dwell, a model-based reinforcement learning framework that turns a pretrained vision-language model into a recurrent latent transition operator. They show that agents can improve their behavior by imagining futures in the model's own latent space, without reconstructing observations. Dwell achieves the strongest overall performance among trainable methods across five interactive visual environments while substantially reducing the compute needed for imagination.

·2 min
Agents

AI agents can train chatbots and improve their appeal with human feedback

RSI Arena tasked ten AI agents with autonomously post-training a base language model over six days with a fixed budget, then improving it using human preference data. Every agent produced a working assistant; the strongest agent, GPT-6 Astra, closed 76% of the gap between the base model and the NVIDIA Instruct model across 12 benchmarks and matched it in human preference when controlling for response length and formatting. In a second stage, eight of ten agents further narrowed the preference gap after reviewing human comparisons of their models.

·3 min