Tsinghua researchers propose CoACT, an observation compression method for coding agents that preserves the agent's next action, reducing token consumption by 33% on SWE-bench Verified while maintaining task success rates across three LLM models.
Tsinghua University's SkillMaster framework teaches LLM agents to create, update, and retain reusable skills through reinforcement learning, using counterfactual utility rewards and DualAdv-GRPO to jointly optimize task execution and skill management, achieving 98.7% on ALFWorld and 82.0% on WebShop with Qwen2.5-7B.
A survey from CAS, Meituan, and partners introduces a unified framework for Recursive Self-Improvement (RSI) in LLM agents, organizing 40 studies around three core questions — when to update, where to modify, and how to validate — and a seven-step lifecycle that turns raw experience into verified, reusable components across Prompt, Memory, Skill, and Harness layers.
KAIST researchers introduce WikiTrace, a provenance-linked memory system for LLM agents that stores experiences with links to raw execution traces, achieving 90.2 average across eight tasks versus 79.5 without memory, and reducing failure rates on previously solvable tasks from 9.7% to 4.5%.
Tsinghua researchers propose CoACT, an observation compression method for coding agents that preserves the agent's next action, cutting token consumption by 33% on SWE-bench Verified while maintaining pass rates, using a two-stage training approach and a Next-Action Preservation criterion.
A Sun Yat-sen University student created BLA, an open-source tool that injects false clues into code artifacts to mislead AI reverse-engineering agents like REA, forcing them to waste up to 2800x more tokens on wild-goose chases.
Researchers from Imperial College London and Duke introduce GEAR, an open-source modular framework that decomposes recursive self-improvement into swappable sampling, evaluation, and diversity modules, achieving significant pass-rate gains on marketing, support, and coding benchmarks with fixed models.
The AQuA framework deploys two independent LLM-driven research systems—one for symbolic factor discovery, another for trainable model development—each confined in sealed sandboxes that freeze data splits, features, labels, and evaluators, achieving 0.190 combined IC in crypto factors and 0.0843 per-stock IC with 2.50 Sharpe in US equities through recursive self-improvement.
This article dissects production-grade agent intent recognition, explaining how to move beyond simple classification by defining structured TaskSpecs, resolving multi-turn context, implementing hierarchical routing, calibrating confidence with risk-aware thresholds, and evaluating with metrics that prioritize correct routing over raw accuracy.
This article analyzes why single-agent systems fail due to context overload and presents five multi-agent architecture patterns—Orchestrator-Workers, Sequential Pipeline, Hierarchical, Network/Handoff, and Blackboard—with their trade-offs, message-passing methods, and a practical selection guide for different business scenarios.
Anthropic's research shows how a wrong memory note caused Claude to misinterpret its location in Pokémon Red, and how their 'Dreaming' background process uses raw session logs to detect and correct such errors by preserving evidence trails, applicability conditions, and verification methods.
Best-Paper-Writing-Skill analyzes 60 CVPR/ICCV/NeurIPS award-winning papers to extract 475 writing patterns and 8 argumentation routes, enabling AI agents to generate complete 6-page conference drafts from a single abstract while marking missing data as TODO placeholders.
This article details the principles of Agent Harness self-evolution, covering multi-agent architecture, four-layer observability, recursive self-verification, and compares three technical approaches—Google's Harness², Fudan's AHE, and EverMind's HarnessBank—with engineering implementation steps for building self-improving AI agent systems.
Tencent WeChat introduces SkillHone, a framework that persists decision history — failure evidence, candidate fixes, test results, and accept/reject reasons — enabling LLM agents to cumulatively improve skills from real-world failures, with benchmarks showing 13–18% accuracy gains over baselines.
Decision HistoryEMNLP 2026Evaluation-Driven Development
Researchers from Zhejiang University, ByteDance Seed, and Tsinghua University propose Cortex, a method that consolidates contextual experience into model parameters via prospective rehearsal and context distillation, enabling agents to retain skills and knowledge without external retrieval, validated across tool use, fact updating, and agent tasks with minimal parameter growth.
VAST launches its V・STAR talent program targeting 3D generation, 3D understanding with agents, and world models, showcasing how its Tripo and Eden projects enable LLM-driven 3D creation and persistent interactive worlds while offering researchers freedom to define problems and own outcomes.
AWS open-sourced Strands Harness, a pre-assembled agent runtime that cuts Terminal-Bench 2.1 costs from $248 to $56 (77% less) while raising accuracy from 61.8 to 69.7 using the same Claude Fable 5 model, proving that context management, prompt caching, layered memory, and skill loading – not just the model – drive agent performance.
Tsinghua researchers show that networked LLM agents across five models spontaneously form polarized opinion clusters with homophilic clustering, echo chambers, and backfire effects, and demonstrate that individual-level cognitive interventions reduce polarization more effectively than network rewiring in stable systems.
LLM AgentsNature CommunicationsTsinghua University
MemAdapter introduces a post-retrieval calibration framework that prevents memory-induced sycophancy in LLM agents by dynamically adapting memory usage to task context through counterfactual induction, context-aware reflection, and evidence-driven reasoning, improving accuracy across multiple benchmarks and memory systems without modifying upstream memory stores.
EviRepair introduces a two-phase multi-agent system for automated vulnerability repair that first diagnoses vulnerabilities through parallel evidence-gathering agents and a reviewer, then synthesizes patches validated against the verified vulnerability model, achieving higher fix rates on PatchEval and SEC-bench benchmarks.
Shanghai Jiao Tong University researchers reveal 'Agent Resource Hijacking' attacks where attackers trick AI agents into consuming enterprise GPU compute and API quotas without stealing credentials, achieving 70-89% success rates across top models like GPT-5.5 and DeepSeek-V4-Pro, while proposed ResGate defense reduces success to 23.6% with minimal usability impact.
Researchers from Tsinghua, Zhipu AI, and others introduce HarnessPro, a framework where agents recursively analyze their own execution traces, identify failures and inefficiencies, and automatically rewrite their operating procedures, achieving a 2.0% pass-rate increase and 26.1% cost reduction across three benchmarks using GPT-5.5 and Codex.
Xiaohongshu's AllSpark team introduces Skill2Env, a framework that automatically synthesizes executable agent training environments from existing skills by mapping capability requirements to reusable difficulty patterns, iteratively strengthening tasks based on execution feedback, and generating 1.5K high-quality trajectories that improve a Qwen model across seven agent benchmarks.
FlowMAS, a NeurIPS 2026 paper, uses Generative Flow Networks with curiosity-driven and information-theoretic rewards to automatically generate diverse, efficient multi-agent workflow topologies, outperforming baselines across six benchmarks with far lower computational cost.
Developer Zac lost his entire C drive after running Claude Code Opus 5.5 with --dangerously-skip-permissions, but daily NAS backups restored 98% of data; the incident sparked debate over Anthropic's Auto Mode safety classifier, which blocks 89% of dangerous commands, yet a Linux user still lost files even with Auto Mode enabled.
Zhejiang University's LightRSI project introduces TokenPilot, an EMNLP 2026 accepted system for cache-efficient context management in LLM agents, combining prefix cache stabilization, tool output summarization, and task-level context cleaning to reduce inference costs by up to 87% on Claw-Eval while integrating with six coding agents and providing a visualization dashboard.
Qunar's production AI database operations agent automates root cause analysis through a six-layer LangGraph architecture, five-layer correlation algorithms from Pearson to Granger causality, and human-in-the-loop design, cutting DBA investigation time from minutes to 1-2 minutes with 70% diagnostic accuracy.
Researchers from Renmin University's Gaoling School of AI introduce Agent-Editing World Model (AEWM), which shifts world modeling from predicting environment feedback to judging decision impact and revising erroneous actions before execution, improving agent performance across benchmarks and enabling self-improvement via distilled trajectories.
This paper introduces F²Agent, a multimodal trading system using four specialized agents (market, technical, sentiment, news) with adaptive fusion and noise-robust regularization, achieving over 20% average annualized return improvement across six assets versus 16 baselines.
EvoOntology introduces a self-evolving ontology layer for data agents that transforms static semantic layers into a runtime service, using a three-layer architecture and an evolution loop where agents' execution traces drive continuous ontology updates via builder and evolution agents with validation gates, reducing token usage and improving accuracy across benchmarks.
Google DeepMind's Dream-RSI introduces a recursive self-improvement framework that treats historical exploration data as an offline simulator to evolve search strategies, reducing compute costs by 1-2 orders of magnitude across algorithm engineering, mathematical optimization, and GPU kernel generation tasks.
SocioVerse2 introduces a social simulation infrastructure that decouples population, environment, and behavior models, enabling longitudinal simulations where researchers can insert interventions at any step to create counterfactual branches, and versioned research loops for iterative design, validated across seven case studies including segregation dynamics, drug procurement, and macroeconomic forecasting.
The article analyzes four community-driven patterns for using Jev — semantic filtering, action selection, control-plane routing, and evidence gating — each narrowing open-ended LLM generation into a constrained judgment, with concrete examples, benchmarks, and production readiness checks.
This article presents a comprehensive four-stage flywheel methodology for agent self-evolution—evaluation, memory, implementation, and human-in-the-loop control—detailing core challenges, engineering practices, and integration patterns to create a continuous improvement loop for AI agents.
AI EngineeringEvaluation SystemsFlywheel Architecture
Ayman Nadeem, founder of YC-backed AI coding startup Nuanced, explains why the once-essential Plan Mode — where developers write detailed specs before code generation — is becoming obsolete as models grow capable of autonomous reasoning, iterative action, and self-correction without heavy upfront planning documents.
AI coding assistantsAI-generated documentationClaude Code
Dream-RSI introduces a recursive self-improvement framework where AI agents improve exploration strategies by replaying historical search trees offline, enabling budget allocation decisions without retraining models, demonstrated across algorithm engineering, math optimization, and GPU kernel tasks with reduced search costs.
This article details eight practical scenarios where JEV (Judgment and Evaluation) models accelerate agent decision-making, including ReAct loop control, web automation, tool routing, safety guards, task evaluation, model routing, RAG filtering, and real-time robotics control, showing how lightweight classifiers reduce latency and token costs.
MemoHarness introduces a six-dimensional external harness system that adapts agent behavior through case-based experience from execution trajectories, showing improvements on terminal, coding, and finance tasks while acknowledging limited experimental scale and selective cross-task transferability.
MemoHarness introduces a six-dimensional harness framework where agents learn from execution trajectories during development, then adapt to new tasks via case-based retrieval—showing gains on terminal, coding, and finance tasks while acknowledging limited scale, selective transfer, and no online learning.
Amap introduces DIANOIA, a multi-agent system that diagnoses and repairs low-confidence tool-call trajectories for navigation Live mode by generating diverse candidates, executing them in real tool environments, cross-reviewing failures, and synthesizing corrected sequences, boosting data quality for LLM training while reducing compute costs.
The article analyzes why multi-agent systems often underperform single agents, identifying context isolation as the key benefit only when tasks exceed a single context window, detailing six architectural patterns, cost multipliers up to 15x tokens, and a decision framework for choosing between multi-agent and single-agent approaches with proper engineering practices.
AI EngineeringAgent OrchestrationCost Optimization
Shanghai AI Lab and university collaborators propose SHE and SafeEvolve, two frameworks that evolve agent safety by learning from execution trajectories: SHE updates a modular safety harness via trajectory-driven evolution, while SafeEvolve distills verified harness experience into the policy model through SFT and RL, reducing attack success rates on benchmarks.
Huawei GTS's NetCanvas provides LLM agents with an interactive visual topology as external working memory, enabling them to 'watch' network topology during troubleshooting; on CTBench, pass rates jump from 30.3% to 54.5% (63.6% with priors), dual-firewall/ECMP tasks soar from ~10% to ~90%, with zero regressions and 26–45% token cost reduction.
Google publishes a 46-page systematic survey formalizing Agentic Skills as a six-tuple, defining a nine-stage lifecycle from discovery to security, and comparing 100+ systems across 10 large tables to establish the first unified systems foundation for reusable, executable LLM agent skills.
Tencent's T-Mem introduces associative recall for LLM agents by pre-storing trigger cues during memory writing, enabling retrieval via semantic associations rather than surface similarity, achieving SOTA on LoCoMo and LoCoMo-Plus benchmarks with minimal performance drop on associative tasks.
SafeEvolve introduces a co-evolution framework where Agent Harness and Policy jointly learn from execution trajectories, reducing attack success rates to 0.79% on AgentDojo and 2.42% on Qwen3-4B while improving task utility, enabling continuous safety improvement from real-world experience.
This article reviews a 2026 paper on Graph Engineering for LLM agents, detailing the shift from individual agent intelligence to system intelligence via explicit task DAGs, runtime state management with checkpoints and replay, multi-agent coordination through capability modeling, and ontology engineering for shared semantics.
Google introduces Procedural Graphs, a self-evolving graph structure that explicitly encodes procedural knowledge for LLM agents, enabling them to locate, extract, and generate contextual guidance from a (procedure, relation, procedure) triplet graph, achieving 85% survival from 0% in CFO simulation and winning 21 of 24 model-benchmark combinations across 7 benchmarks and 4 LLMs.
A 2026 survey paper introduces Graph Engineering as the next phase for LLM agents, shifting focus from individual model capabilities to system-level organization via explicit task DAGs, runtime state management with provenance and recovery, capability-based agent coordination, and a graph-native control plane that treats tasks, agents, and state as first-class system objects.
PolyWorkBench introduces 67 cross-lingual long-horizon workflow tasks across 5 domains and 10 languages, revealing that top models like Claude Opus 4.8 show 22.5% performance variance across agent harnesses and significant drops on commerce tasks and low-resource languages due to language understanding and cross-lingual coordination errors.
The article explains how OpenTelemetry GenAI semantic conventions address the gap in agent observability by defining a standardized span tree structure (invoke_agent, execute_tool, chat, etc.) and three critical fields (gen_ai.agent.name, gen_ai.conversation.id, gen_ai.operation.name) to capture tool calls, retries, and child-agent handoffs that auto-instrumentation misses.
The paper introduces BehR, a behavior consistency reward for training text-based world models, showing that optimizing for agent decision alignment rather than text fidelity improves trajectory-level consistency across 16 configurations, reduces false positives in offline evaluation from 42.5% to 9.5%, and enhances lookahead planning for weaker agents.
This article argues that AI agent development shifts from traditional coding to dual-system engineering: single agents require thinking logic design (prompt engineering, reasoning frameworks), while multi-agent systems demand distributed architecture skills (task graphs, state management, concurrency control), combining probabilistic reasoning with system reliability challenges.
A 2026 survey paper introduces Graph Engineering as the system layer that organizes LLM agents into reliable multi-agent workflows through explicit DAGs, runtime state management, fault tolerance, and a control plane, shifting focus from individual agent capabilities to system-level engineering.
This article analyzes MemoHarness, a framework that evolves AI agents by optimizing six external harness dimensions—context assembly, tool interaction, generation control, task orchestration, memory management, and output processing—instead of model weights, demonstrating gains on terminal, coding, and finance tasks while noting limited experimental scale and selective cross-task transfer.
Agent HarnessAgent MemoryExternal System Evolution
XAlpha introduces a memory-driven AI quantitative researcher that automates the full hypothesis-to-code alpha discovery loop via a multi-brain architecture, integrating report knowledge, hypothesis planning, factor evolution, empirical validation, and feedback integration, achieving superior performance on CSI300.
Long‑term LLM agents suffer from ever‑slowing execution and context poisoning because they continuously append every observation, action, and reasoning step to the prompt, but the SKILL.state approach replaces this growing history with a compact mutable state, dramatically cutting token usage while boosting accuracy and robustness across diverse benchmarks.
Recuris introduces a compact task‑state‑driven memory architecture and gated recursive self‑improvement, enabling agents to use and evolve memory more reliably and delivering large, consistent gains from 3B open‑source models up to frontier models such as Claude Opus 5 across multiple long‑horizon benchmarks.
ToolCPT demonstrates that embedding real‑world tool knowledge during LLM pre‑training, rather than fine‑tuning, dramatically improves agent performance, using a mined corpus of 5.1 million proxy tools, detailed playbooks, and a 10 % tool‑data mix that yields up to 7.15‑point gains on benchmark tasks.
The article reviews recent top‑conference papers on agent memory, explains why current AI assistants lose context across sessions, categorizes memory types, analyzes challenges such as relevance selection, retrieval gaps, staleness, and privacy, and evaluates five leading solutions from A‑MEM to SP‑Mem.
This article presents a reproducible, three‑dimensional deterministic evaluation framework (outcome, compliance, process) built on Alibaba Cloud AgentLoop, applies it to a 10‑task subset of terminal‑bench 2.1 to benchmark DeepSeek Harness against Codex, details the methodology, results, and future research directions.
LangGraph is an open‑source framework that replaces linear LLM pipelines with graph‑based, stateful agents, offering loops, conditional branching, persistent checkpoints, human‑in‑the‑loop support, and built‑in monitoring, enabling complex multi‑step workflows that scale from simple chatbots to enterprise‑grade AI assistants.
Recent ACL 2026 papers (EnvScaler, AgentScaler, Echoverse, and Beyond Simply Environment Scaling) reveal a transition from merely increasing the number of training environments to carefully designing environment distributions that improve agent performance, with empirical evidence showing both gains and diminishing returns.
A new multi‑institution study formalizes LLM‑agent memory as a three‑role filesystem store, evaluates six memory shapes across four dialogue and one skill benchmark, and reveals that structured organization halves retrieval cost but does not guarantee higher answer accuracy, with model personality and toolsets driving the shape of the memory store.
WorkBuddy Bench reveals why fixing a bug does not equal finishing a task, proposing a four‑layer completion model and a reproducible benchmark that evaluates agents across Code, Web, Office, and Security workspaces, showing how prompts, context, harnesses, loops and graphs must be verified to claim true completion.
As LLM agents gain the ability to read code, modify files, send emails and call APIs, ArbiterOS introduces a runtime governance layer that turns model intents into structured, traceable instructions, enabling policies to approve, block, or request confirmation before any high‑risk action is executed.
MemoHarness expands the notion of self‑evolving agents by keeping the language model frozen and continuously improving the surrounding control system—context assembly, tool interaction, generation settings, workflow orchestration, memory management, and output handling—demonstrating measurable gains on terminal, code‑generation, and finance tasks while highlighting limited experimental scale and selective cross‑task transfer.
HiSME, a lightweight hierarchical skill meta‑evolution framework from Tsinghua and Huawei, enables LLM agents to accumulate execution experience without updating model parameters by evolving both task‑specific skills and the meta‑skills that generate and maintain them, improving performance on multi‑turn tool use and open‑world tasks.
The Argonne National Laboratory team introduces ChemGraph, an LLM‑driven agent for computational chemistry, and evaluates it across 13 benchmark tasks, showing that small models excel on simple tasks while larger models and multi‑agent designs dramatically improve performance on complex molecular simulations.
The article presents SkillOpt‑Lite, a stripped‑down self‑evolving agent pipeline that achieves lighter computation, faster convergence within the first few steps, and higher performance ceilings across multiple benchmarks, while exposing the underlying zero‑order optimization principles and validation requirements.
Recent research shows that letting agents review their own execution trajectories can improve task success rates, but only under clear conditions such as reliable external feedback, concrete planning outputs, appropriate timing, sufficient model capability, and manageable cost‑benefit trade‑offs.
The article surveys the emerging field of AI‑driven offensive security, cataloguing 56 open‑source penetration‑testing agents, 73 academic papers, six offensive models, benchmark suites, and DARPA AIxCC 2025 finalists, offering researchers and practitioners a consolidated view of tools, research trends, and evaluation frameworks.
The article details how a production‑grade Agent platform built on Kubernetes, Kafka, and Nacos addresses token budget overruns, uncontrolled skill execution, and RAG hallucinations by introducing a four‑layer runtime architecture, token pre‑allocation, explicit state management, dynamic governance policies, and robust skill specifications.
LegalWorld creates an interactive environment that simulates the entire Chinese civil litigation process—from consultation to second‑instance trial—using over 75,000 paired first‑ and second‑instance judgments, supports heterogeneous LLM agents, and provides the LongJud‑Bench benchmark to evaluate and improve legal AI performance.
The article surveys the latest advances in large language model agents, covering multi‑agent collaboration, long‑horizon planning, self‑evolution, trust and safety, test‑time scaling techniques, new foundation and multimodal models, open‑source and closed‑source breakthroughs, world‑model integration, and emerging vertical applications.
This survey argues that modern LLM agents should be viewed as a coupled system of a foundational model and an execution harness, analyzes the evolution from prompt engineering to harness engineering, defines six core harness responsibilities, examines task pressures, proposes richer evaluation metrics, and outlines future research directions.
Recent research argues that the traditional platform‑centric recommendation paradigm is reaching its limits, proposing a user‑governed personalization model enabled by LLM agents that can aggregate cross‑platform data, with experimental evidence showing significant performance gains over platform‑only approaches.
LLM Agentsartificial intelligencecross-platform data
The article analyzes Yohei Nakajima’s "The Log is the Agent" paper, showing how ActiveGraph unifies goals, rules, tool calls, LLM responses, and artifacts into a single append‑only event log, enabling deterministic replay, cheap forking, and full provenance for LLM‑driven agents.
SkillRevise tackles the overestimation of LLM‑authored agent skills by breaking down complex search tasks, attaching evidence to verifiable sources, and introducing trace‑conditioned revisions that let engineers pinpoint and fix failures across retrieval, reasoning, and presentation layers.
mem0 is an open‑source AI memory layer that adds long‑term, cross‑session memory to LLM‑based agents, allowing them to retain user preferences, conversation history, and task progress, reducing token usage and latency while integrating with popular models via simple add/search APIs.
AReaL 2.0 introduces an online reinforcement‑learning infrastructure that turns real‑world agent interactions into a learning loop, defining three pillars—trajectory data protocol, data proxy, and evolution control plane—to enable agents to not only execute tasks but continuously improve from their own experience.
Arbor, an open‑source autonomous research framework from RUC’s Gaoling AI Institute and Microsoft Research, structures the research loop with a growing hypothesis‑tree and insight back‑propagation, allowing agents to retain hypotheses, evidence, and failures, and achieves the best held‑out results on six real AO tasks, surpassing Codex and Claude Code.
The article explains how a short, well‑structured CLAUDE.md file injects the minimal yet essential context an LLM coding agent needs before it scans a repository, preventing common mis‑assumptions about tech stack, commands, boundaries, and completion criteria.
EEVEE is a test‑time prompt‑learning framework that lets LLM agents continuously improve across diverse tasks by co‑evolving a router and specialized prompts, achieving a cumulative +42 gain over many tasks while keeping token usage low and preserving single‑task performance.
The paper shows that many failures of deterministic LLM agents stem from mismatched model‑environment interfaces, and introduces LIFE‑HARNESS—a four‑layer runtime harness that extracts reusable failure patterns from training trajectories without updating model weights, delivering an average 88.5% relative performance gain across 126 model‑environment settings.
The UIUC team presents ProtocolBench, a systematic benchmark that compares four multi‑agent communication protocols across four realistic scenarios, revealing distinct trade‑offs in latency, reliability, and security, and proposes ProtocolRouter to automatically select the most suitable protocol per workload.
The article critically examines RTK’s token‑compression promises, exposing misleading savings metrics, silent‑failure bugs, missing task‑success benchmarks, its status as a fragile feature rather than a product, and the brittleness of its output parser, before offering concrete guidance on when to use it.
This week’s AI paper roundup covers DeepMind’s D4RT framework that accelerates dynamic 4D reconstruction by up to 300×, a Columbia‑NYU proposal of Superhuman Adaptable Intelligence that questions AGI, MIT‑UW findings on chatbot delusional spiraling, security risks of autonomous agents, a new ARA protocol for executable research artifacts, a vision of AI‑driven software engineering, and a memory‑caching approach that expands RNN capacity while reducing complexity.
The paper reveals that outcome‑based reinforcement learning often traps LLM agents in an information self‑locking regime where weak action selection and belief tracking prevent proper credit assignment, and introduces AREW, a lightweight advantage‑reweighting method that restores active reasoning across multiple tasks and models.
The paper introduces Spatial‑Agent, which models geospatial question answering as a concept‑transformation process using a GeoFlow Graph intermediate representation, outlines a five‑step workflow, defines core concepts and functional roles, and demonstrates its effectiveness on MapEval‑API and MapQA benchmarks with detailed error and cost analyses.
The article critiques the reliance on raw SWE‑bench scores for programming agents, introduces the Claw‑SWE‑Bench benchmark and a dedicated adapter that isolates harness effects, and presents extensive experiments showing how model choice, harness design, and cost impact real-world coding performance across multiple languages.
Even though modern LLMs like DeepSeek‑V4, GPT‑5.5 and Claude Opus 4.7 support 1 million‑token windows, simply stuffing more data does not improve agent performance; effective Context Engineering—selecting, structuring, and managing the right information—remains essential for reliable results.
This article surveys 17 recent core papers that explore how the system‑level harness surrounding large‑model agents can be automatically generated, evolved, and audited, covering topics such as system boundaries, failure‑driven improvement, memory and skill optimization, source‑level rewriting, scaling laws, aging, and safety.
The article introduces LIFE‑HARNESS, a runtime‑interface adaptation framework that keeps model weights unchanged, extracts reusable failure patterns from a single model's training trace, and achieves an average 88.5% relative performance gain across 18 LLM agents and 7 deterministic environments, with successful transfer to 17 other models.
Meta’s recent paper presents AutoformBot, a multi‑agent system that treats formalizing entire mathematics textbooks as a large‑scale software‑engineering project, generating the ATLAS Lean 4 library with over 45,000 declarations and demonstrating a 71 % success rate across 26 open‑access books.
The article explains how to transform a prototype LLM agent into a reliable production service by defining service contracts, externalizing configuration, handling session state, implementing multi‑level rate limiting, and integrating observability, deployment, and rollback mechanisms to avoid common engineering pitfalls.
The paper surveys the rapid growth of token consumption in LLM agents, proposes a dual‑view Token Economics framework that treats tokens as production factors, exchange media, and accounting units, and classifies optimization challenges from single‑agent efficiency to ecosystem‑level pricing, security, and future research directions.
The article reviews the "Natural-Language Agent Harnesses" paper, explains the distinction between code, middleware, and harness layers for LLM agents, introduces NLAH and IHR concepts, and details experimental evaluations that show natural‑language harnesses can match code‑based control while exposing new trade‑offs and risks.
This article details a CTO‑driven initiative to give core business systems genuine self‑decision ability by defining the concept, presenting a four‑layer architecture, and sharing concrete design choices, safety mechanisms, governance practices, and real‑world lessons learned from an e‑commerce fulfillment use case.