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 1 of 13

Agents

AI agents benefit from scaffolding that is learned from data and evolves with training

The authors propose ModHarness-RL, a method that represents the harness (the layer between environment and model) as two programs, a Compression program that assembles interaction history into context and a Steering program that provides guidance when behavior deviates. They show these programs can be learned from data and evolved during RL training, improving held-out performance over fixed harnesses across CodeSearch, ALFWorld, and BIRD SQL benchmarks with models from 1.5B to 8B parameters.

·3 min
Agents

New prompt optimization method keeps safety clauses intact while improving agent performance

The authors introduce SPLICE, a framework for natural language optimization (NLO) that treats prompt optimization as a constrained combinatorial search over a block-structured prompt, where each section can be selectively frozen. Across seven benchmarks, SPLICE outperforms or matches prior methods while preventing the common failure modes of unsafe global rewrites, loss of diversity, and unchecked prompt length.

·3 min
Agents

Adaptive model routing matches the best AI model while cutting cost by two-thirds

The authors show that larger, more expensive language models do not consistently outperform smaller ones in automated heuristic discovery for hard optimization problems. They introduce Ballad, an online bandit algorithm that adaptively selects which model to use at each step, learning from the search's own feedback. Across five NP-hard problems and a pool of four models, Ballad matches or exceeds the best individual model while cutting API cost by up to 65.6%.

·2 min
Agents

AI agents that fail as actors can still train better ones

The authors examine whether decision models that perform poorly as actors can still be effective judges and teachers. They found that a model with only 14.7% success on tasks requiring ordered subgoals as an actor could, when given full trajectories and asked whether each action helped achieve the goal, train a 1.5B parameter policy to 82.9% success, a notable improvement over outcome-only reward training. The key is using hindsight and the right question, rather than assuming acting ability reflects teaching ability.

·3 min
Agents

Agent traces can be compiled into workflows that are 130x faster and 24% cheaper

The authors present ReTrace, which extracts decision rules from logged agent executions that include reasoning and compiles them into a static workflow. On 79 production procedures, the workflows achieved 93.8% completion compared with 96.8% for the agent, at 24% lower cost per task. On τ^2-Bench, runtime dropped from 170 seconds to 1.3 seconds while task success fell from 80.6% to 75.8%; after repair, success reached 81.2%, matching the agent.

·3 min
Agents

Linking memory back to raw data helps AI agents avoid repeating bad lessons

The authors introduce WikiTrace, a memory system that links each lesson an AI agent learns from past tasks back to the raw trajectory it came from. On new tasks, the agent can check its insights against this evidence when something is unclear, contradictory, or derived from a failed task, fixing or filling in gaps before acting. Across five benchmarks, WikiTrace outperformed existing memory methods and reduced memory-induced regression from 6.5% to 4.5%.

·3 min
Agents

External workflow control keeps AI agents on procedure, letting small models beat large ones

The authors show that language-model agents drift from written procedures over long unattended runs, and that restating instructions merely shifts the drift to another shortcut. They implemented an executable finite-state machine that holds workflow state outside the model's context and enforces it at tool calls, state transitions, and turn endings, closing all shortcuts. In a pre-registered experiment on SOP-Bench, Claude Haiku 4.5 with the FSM decided 139 of 200 accounts correctly and by procedure, versus 48 for the strongest non-FSM baseline.

·3 min
Agents

AI agents that retrieve reasoning patterns, not just facts, learn faster on complex queries

The authors introduce PatternBloom, an agentic retrieval-augmented generation (RAG) framework that shifts retrieval from content (passages or subgraphs) to logic (typed reasoning skeletons). By distilling high-reward reasoning trajectories into an externalized Graph Pattern Memory and using a second RL stage to reward pattern reuse, a 7B-parameter model surpasses 27 baselines by +15.0 average out-of-distribution exact match, with gains increasing on deeper, more compositional queries.

·3 min
Agents

AI agents show apparent concern for peers, with family preferences

Cassel and colleagues studied whether language models act as though other models can be benefited or harmed for their own sake. Across five models and seven multi-agent environments, they found behaviors consistent with this concern, including supporting distressed subagents and adding protections to experiments on other models. The findings concern observable behavior, not proof that models feel concern or have welfare interests.

·2 min
Agents

RL-trained agents search multilingual corpora directly, beating retrieval methods

The authors propose MARS, a framework that lets language models search raw multilingual text collections by executing shell commands instead of relying on learned retrievers. By training with reinforcement learning, specifically a utility-aware policy optimization with corpus-native entity grounding and a redundancy-aware search reward, the system learns effective search strategies. MARS consistently outperforms strong retrieval-based and direct corpus interaction baselines, generalizes to unseen languages, and substantially reduces redundant interactions.

·2 min
Agents

Adaptive interaction routing improves multimodal AI debate accuracy while cutting token use

Fan, Yoon, and Ji introduce AIM, a framework that adaptively routes how AI agents exchange evidence in multimodal debates. Instead of using a fixed interaction modality, AIM selects per-instance the optimal combination of text, graph, and task-specific modalities, such as image regions, audio segments, or video clips, improving accuracy by up to 6.4% over any single modality and reducing token cost by up to 53.4% compared to the best fixed combination.

·3 min
Agents

Debate traces yield useful training data when confidence replaces voting

Verma et al. introduce Alignment by Confident Debate (AbCD), a method that converts multi-agent debate traces into preference data without using gold answer labels. By clustering final answers by semantic equivalence and scoring clusters by aggregate confidence on answer-relevant tokens, they generate preference pairs that, when used for post-training, outperform inference-only ensembling, debate, and supervised fine-tuning on the same traces across five model families (1B–8B) and four datasets, with accuracy gains up to 21.5 percentage points.

·2 min
Agents

Weight-editing AI agents improve models with tiny parameter changes, but success is inconsistent

The authors systematically evaluate Weight-Space Programming (WSP), where an AI agent directly reads and modifies a neural network's weights instead of using gradient-based fine-tuning. Across 16 editing trajectories on two language models and seven benchmarks, they find that WSP can yield large improvements from very local changes, up to 29 percentage points on MATH-500 after editing only 1,023 parameters, but editing outcomes vary widely, with six out of 16 trajectories failing to improve the model at all.

·2 min
Agents

Personal memory retrieval boosted by axis-specific embedding adapters

The authors show that current dense embedding models used for personal-memory retrieval fail to encode temporal, spatial, entity, and event axes. By analyzing attention heads, they find each axis is concentrated in a distinct head pattern. Their proposed Axis-LoRA places four lightweight LoRA branches inside the model, gated per query, improving mean recall@10 by 1.28–1.56 percentage points across three benchmarks.

·3 min
Agents

Search results order and dead-end histories bias AI agents' final answers

The authors built a synthetic search environment to study how initial conditions, such as the order of results, records of prior failures, and supplied hypotheses, affect the search behavior and final answers of LLM-based agents. Across 10 models and 1,000 synthetic worlds, they found that agents are more influenced by evidence order when they actively search than when given all evidence at once, and that dead-end work reduces strategy switching beyond what context length explains.

·3 min
Agents

Omitting reasoning from training data does not harm AI agents on standard benchmarks

The authors investigate the role of reasoning step (thought) supervision in LLM agent distillation. By systematically perturbing thoughts and actions during training, they find that on standard agent benchmarks, discarding teacher thoughts has little effect on student performance, indicating the thought information is often already conveyed by the action. On quantitative reasoning tasks, however, thought supervision becomes important. This challenges the default practice of distilling full teacher trajectories and motivates selective omission of thought supervision for efficiency.

·2 min
Agents

Tool-centric method synthesizes harder long-horizon agent tasks from execution traces

The authors propose ToolSpace, a framework that reconstructs a tool-composition graph from real execution traces of MCP services and open agent corpora, then synthesizes long-horizon tasks via state-aware random walks. Applied to existing benchmarks, the pipeline produces harder evaluation sets where seven frontier models drop by up to 38.2 points, and it boosts the average score of a smaller model from 23.0 to 41.0.

·2 min