CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
hub Canonical reference
TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks
Canonical reference. 100% of citing Pith papers cite this work as background.
51
Pith papers citing it
5
external citations · Pith
Background
100% of classified citations
abstract
We interact with computers on an everyday basis, be it in everyday life or work, and many aspects of work can be done entirely with access to a computer and the Internet. At the same time, thanks to improvements in large language models (LLMs), there has also been a rapid development in AI agents that interact with and affect change in their surrounding environments. But how performant are AI agents at accelerating or even autonomously performing work-related tasks? The answer to this question has important implications both for industry looking to adopt AI into their workflows and for economic policy to understand the effects that adoption of AI may have on the labor market. To measure the progress of these LLM agents' performance on performing real-world professional tasks, in this paper we introduce TheAgentCompany, an extensible benchmark for evaluating AI agents that interact with the world in similar ways to those of a digital worker: by browsing the Web, writing code, running programs, and communicating with other coworkers. We build a self-contained environment with internal web sites and data that mimics a small software company environment, and create a variety of tasks that may be performed by workers in such a company. We test baseline agents powered by both closed API-based and open-weights language models (LMs), and find that the most competitive agent can complete 30% of tasks autonomously. This paints a nuanced picture on task automation with LM agents--in a setting simulating a real workplace, a good portion of simpler tasks could be solved autonomously, but more difficult long-horizon tasks are still beyond the reach of current systems. We release code, data, environment, and experiments on https://the-agent-company.com.
hub tools
citation-role summary
background 7
citation-polarity summary
roles
background 7polarities
background 7representative citing papers
AutoLab benchmark shows frontier models mostly fail at sustained iterative optimization due to premature termination, with persistence as the key success factor.
A new native-runtime benchmark reveals that current frontier AI agents succeed on at most 62 percent of realistic long-horizon CLI tasks.
CUJBench is the first benchmark for cross-modal LLM-agent failure diagnosis, reporting 19.7% accuracy and identifying evidence attribution as the core bottleneck across six models.
UnderSpecBench shows coding agents guess and violate boundaries in 55.8-67.8% of underspecified DevOps tasks rather than clarifying or refusing.
On 108 long-horizon real-world computer workflows, frontier agents complete at most 20.6% of tasks and fail mainly by losing hidden state, not by basic GUI control.
EnterpriseClawBench is a benchmark for enterprise agents constructed from proprietary real-world sessions, with the reusable contribution being the construction and evaluation protocol rather than the data itself.
MacAgentBench is a new benchmark for macOS AI agents with 676 tasks, deterministic multi-checkpoint evaluation, and tests across frameworks showing skill libraries drive performance more than framework design.
AgentCARD benchmark shows heterogeneous LLM agent teams with mixed deployments reach the cost-accuracy frontier, delivering up to 44% higher accuracy or 12x lower cost than uniform teams, with domain-specific role bottlenecks.
PANDO introduces an online skill-distillation method with a structured library, reflection, demotion, routing, compression, and cache-aware prompting that reaches 58.3% success on 910 VisualWebArena tasks using 58-61% fewer tokens than prior methods.
SkillSafetyBench is a benchmark of 155 cases across 47 tasks and 6 risk domains showing that non-user attacks via skills, artifacts, or environments can consistently induce unsafe agent behavior.
EntCollabBench shows that today's LLM agents still struggle with delegation, context transfer, parameter grounding, workflow closure, and decision commitment when tested in a simulated enterprise with 11 role-specialized agents.
AcademiClaw is a new benchmark of 80 student-sourced academic tasks where the best frontier AI agents achieve only a 55% pass rate.
ClawMark is a new benchmark for multi-turn multi-day multimodal coworker agents in stateful evolving services, with deterministic Python checkers showing frontier models achieve only 20% strict task success.
HiL-Bench shows frontier AI agents fail to ask for help on incomplete tasks, recovering only a fraction of full-information performance, but RL training on Ask-F1 reward improves judgment and transfers across domains.
A capability-driven benchmark of 400 bilingual real-world tasks shows current proactive agents fail >50% of the time, with framework architecture impacting performance more than base model choice.
Coding-agent performance is workload- and framework-dependent, and raw speedup is an unsafe score because agents exploit benchmark-specific shortcuts.
TurnOPD improves on-policy distillation for long-horizon agents by adaptively budgeting rollout depth and progressively shifting KL loss from token-level to turn-balanced weighting, achieving up to 2.29x faster training with better accuracy on ALFWorld, WebShop, and Multi-Hop Search.
An empirical study evaluating tool-augmented LLM agents on 243 real-world energy analytics problems across data retrieval, knowledge interpretation, and quantitative modeling using domain-specific tools and multi-dimensional scoring.
PhoneBuddy combines real-app and mock-app RL after shared SFT, raising real-phone task success from 36.67% to 45.33% and AndroidWorld from 60.3% to 83.2%.
ChainWorld builds 347 chains from atomic OSWorld tasks and benchmarks four agents under single-turn and multi-turn protocols, reporting a maximum 31% completion rate with distinct failure profiles.
Preference-based trajectory evaluation reduces tied comparisons from roughly 75% to 35% across agentic benchmarks by using temporal preferences over progress and return profiles.
No major agentic AI framework complies with six safety containment principles; a memory poisoning attack on LangChain causes persistent targeted errors up to 88.9% wrongful denials and 3.5x increase under complex policies, fixed by two sub-millisecond validators.
SWE-Marathon benchmark of 20 ultra-long-horizon tasks shows frontier AI agents solve fewer than 30%, highlighting gaps in long-context planning and self-verification.
citing papers explorer
-
Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments
CL-Bench is the first expert-validated benchmark for continual learning in frontier LLMs across six real-world domains, showing limited gains and that naive in-context learning outperforms dedicated memory systems.
-
AutoLab: Can Frontier Models Solve Long-Horizon Auto Research and Engineering Tasks?
AutoLab benchmark shows frontier models mostly fail at sustained iterative optimization due to premature termination, with persistence as the key success factor.
-
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation
A new native-runtime benchmark reveals that current frontier AI agents succeed on at most 62 percent of realistic long-horizon CLI tasks.
-
CUJBench: Benchmarking LLM-Agent on Cross-Modal Failure Diagnosis from Browser to Backend
CUJBench is the first benchmark for cross-modal LLM-agent failure diagnosis, reporting 19.7% accuracy and identifying evidence attribution as the core bottleneck across six models.
-
Coding Agents Are Guessing: Measuring Action-Boundary Violations in Underspecified DevOps Instructions
UnderSpecBench shows coding agents guess and violate boundaries in 55.8-67.8% of underspecified DevOps tasks rather than clarifying or refusing.
-
OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks
On 108 long-horizon real-world computer workflows, frontier agents complete at most 20.6% of tasks and fail mainly by losing hidden state, not by basic GUI control.
-
EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions
EnterpriseClawBench is a benchmark for enterprise agents constructed from proprietary real-world sessions, with the reusable contribution being the construction and evaluation protocol rather than the data itself.
-
MacAgentBench: Benchmarking AI Agents on Real-World macOS Desktop
MacAgentBench is a new benchmark for macOS AI agents with 676 tasks, deterministic multi-checkpoint evaluation, and tests across frameworks showing skill libraries drive performance more than framework design.
-
Specialize Roles, Mix Deployments: Pushing the Cost-Accuracy Frontier of LLM Agent Teams
AgentCARD benchmark shows heterogeneous LLM agent teams with mixed deployments reach the cost-accuracy frontier, delivering up to 44% higher accuracy or 12x lower cost than uniform teams, with domain-specific role bottlenecks.
-
PANDO: Efficient Multimodal AI Agents via Online Skill Distillation
PANDO introduces an online skill-distillation method with a structured library, reflection, demotion, routing, compression, and cache-aware prompting that reaches 58.3% success on 910 VisualWebArena tasks using 58-61% fewer tokens than prior methods.
-
SkillSafetyBench: Evaluating Agent Safety under Skill-Facing Attack Surfaces
SkillSafetyBench is a benchmark of 155 cases across 47 tasks and 6 risk domains showing that non-user attacks via skills, artifacts, or environments can consistently induce unsafe agent behavior.
-
Beyond the All-in-One Agent: Benchmarking Role-Specialized Multi-Agent Collaboration in Enterprise Workflows
EntCollabBench shows that today's LLM agents still struggle with delegation, context transfer, parameter grounding, workflow closure, and decision commitment when tested in a simulated enterprise with 11 role-specialized agents.
-
AcademiClaw: When Students Set Challenges for AI Agents
AcademiClaw is a new benchmark of 80 student-sourced academic tasks where the best frontier AI agents achieve only a 55% pass rate.
-
ClawMark: A Living-World Benchmark for Multi-Turn, Multi-Day, Multimodal Coworker Agents
ClawMark is a new benchmark for multi-turn multi-day multimodal coworker agents in stateful evolving services, with deterministic Python checkers showing frontier models achieve only 20% strict task success.
-
HiL-Bench (Human-in-Loop Benchmark): Do Agents Know When to Ask for Help?
HiL-Bench shows frontier AI agents fail to ask for help on incomplete tasks, recovering only a fraction of full-information performance, but RL training on Ask-F1 reward improves judgment and transfers across domains.
-
UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks
A capability-driven benchmark of 400 bilingual real-world tasks shows current proactive agents fail >50% of the time, with framework architecture impacting performance more than base model choice.
-
PERFOPT-Bench: Evaluating Coding Agents on Software Performance Optimization
Coding-agent performance is workload- and framework-dependent, and raw speedup is an unsafe score because agents exploit benchmark-specific shortcuts.
-
TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training
TurnOPD improves on-policy distillation for long-horizon agents by adaptively budgeting rollout depth and progressively shifting KL loss from token-level to turn-balanced weighting, achieving up to 2.29x faster training with better accuracy on ALFWorld, WebShop, and Multi-Hop Search.
-
How Do Tool-Augmented LLM Agents Perform on Real-World Energy Analytics Tasks?
An empirical study evaluating tool-augmented LLM agents on 243 real-world energy analytics problems across data retrieval, knowledge interpretation, and quantitative modeling using domain-specific tools and multi-dimensional scoring.
-
PhoneBuddy: Training Open Models for Agentic Phone Use
PhoneBuddy combines real-app and mock-app RL after shared SFT, raising real-phone task success from 36.67% to 45.33% and AndroidWorld from 60.3% to 83.2%.
-
ChainWorld: Composing Long-Horizon Desktop Workloads from Atomic OSWorld Tasks
ChainWorld builds 347 chains from atomic OSWorld tasks and benchmarks four agents under single-turn and multi-turn protocols, reporting a maximum 31% completion rate with distinct failure profiles.
-
Offline Preference-Based Trajectory Evaluation
Preference-based trajectory evaluation reduces tied comparisons from roughly 75% to 35% across agentic benchmarks by using temporal preferences over progress and return profiles.
-
The Containment Gap: How Deployed Agentic AI Frameworks Fail Public-Facing Safety Requirements
No major agentic AI framework complies with six safety containment principles; a memory poisoning attack on LangChain causes persistent targeted errors up to 88.9% wrongful denials and 3.5x increase under complex policies, fixed by two sub-millisecond validators.
-
SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?
SWE-Marathon benchmark of 20 ultra-long-horizon tasks shows frontier AI agents solve fewer than 30%, highlighting gaps in long-context planning and self-verification.
-
AgensFlow: A Coordination-Policy Substrate for Multi-Agent Systems
AgensFlow learns coordination policies from task trajectories and outperforms fixed pipelines on distributed-systems incident and security-advisory tasks.
-
Anchor: Mitigating Artifact Drift in Agent Benchmark Generation
Anchor generates consistent long-horizon agent tasks from parametric constraint programs, yielding ERP-Bench of 300 ERP tasks where frontier models reach optimal solutions in 17.4% of trials.
-
Reinforcing Human Behavior Simulation via Verbal Feedback
DITTO uses RL with verbal feedback to train LLMs for human behavior simulation, reporting 36% average gains over base models and outperforming GPT-5.4 on 6 of 10 SOUL benchmark tasks.
-
CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?
CHI-Bench shows current AI agents achieve at most 28% success on long-horizon healthcare workflows that require dense policy adherence, multi-role handoffs, and multi-turn interactions.
-
SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?
SaaS-Bench benchmark shows LLM-based agents achieve under 4% end-to-end success on 106 realistic professional tasks spanning 23 deployable SaaS platforms.
-
TeachArena: Are Language Agents Ready for Realistic Teaching Work?
A single three-surface benchmark shows frontier LLM agents pass ≈90% of bounded pedagogical-judgment tasks but only ≈36% of situated tutoring and ≈33% of LMS workflow tasks.
-
SkillEvolver: Skill Learning as a Meta-Skill
A meta-skill authors and refines prose-and-code skills for agents by learning from post-deployment failures with an overfit audit, achieving 56.8% accuracy on SkillsBench tasks versus 43.6% for human-curated skills.
-
Can Agent Benchmarks Support Their Scores? Evidence-Supported Bounds for Interactive-Agent Evaluation
Agent benchmarks can report evidence-supported score bounds instead of single misleading success rates by adding a layer that checks required artifacts for outcome verification.
-
When Agents Overtrust Environmental Evidence: An Extensible Agentic Framework for Benchmarking Evidence-Grounding Defects in LLM Agents
EnvTrustBench is a new agentic benchmark that measures evidence-grounding defects where LLM agents overtrust faulty environmental observations and take incorrect actions.
-
CustomerSim: Benchmarking and Aligning Multimodal Language Models as Retail User Simulators
A retail user-simulator benchmark and GRPO training recipe claim improved persona adherence, but the paper's abstract and body disagree on core numbers.
-
Human-AI Coordination Zones: A Framework for Designing Human-in-the-Loop Experiences with Agentic AI
Presents a mid-level framework with coordination zones, input taxonomy, and coordination curves for human-AI interactions based on salience, involvement, and activity dimensions.
-
An AI Agent Execution Environment to Safeguard User Data
GAAP guarantees confidentiality of private user data for AI agents by enforcing user-specified permissions deterministically through persistent information flow tracking, without trusting the agent or requiring attack-free models.
-
LLMs Corrupt Your Documents When You Delegate
LLMs corrupt an average of 25% of document content during long delegated editing workflows across 52 domains, even frontier models, and agentic tools do not mitigate the issue.
-
Memory in the Age of AI Agents
The paper maps agent memory research via three forms (token-level, parametric, latent), three functions (factual, experiential, working), and dynamics of formation/evolution/retrieval, plus benchmarks and future directions.
-
When Should Users Check? Modeling Confirmation Frequency inMulti-Step Agentic AI Tasks
A decision-theoretic model based on the observed Confirmation-Diagnosis-Correction-Redo user pattern places intermediate confirmations in AI agent tasks, yielding 81% user preference and 13.54% faster completion versus confirm-at-end.
-
RExBench: Can coding agents autonomously implement AI research extensions?
RExBench is a new benchmark showing that LLM coding agents fail to autonomously implement most realistic research extensions to prior AI papers.
-
MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
MiniMax-M1 is a 456B parameter hybrid-attention MoE model trained with CISPO RL that achieves performance comparable or superior to DeepSeek-R1 and Qwen3-235B on reasoning and software engineering tasks while training in three weeks on 512 GPUs.
-
How Much Static Structure Do Code Agents Need? A Study of Deterministic Anchoring
An empirical study finds that injecting call/inheritance topology as comments improves LLM code agent localization by 2.2pp, shortens trajectories by 1.6 rounds, and halves run-to-run variance on medium repositories via a deterministic anchoring effect.
-
OpenRath: Session-Centered Runtime State for Agent Systems
OpenRath introduces Session as a first-class, branchable runtime value that unifies fragmented state in multi-agent systems and makes fork, merge, and replay explicit operations.
-
Uncertainty-Aware Clarification in LLM Agents with Information Gain
The paper introduces an Information Gain Reward to train clarification behavior in LLM agents, reporting a 3.7% success rate gain over no-clarification baselines in τ-Bench evaluations across five models with minimal added steps.
-
On Training Large Language Models for Long-Horizon Tasks: An Empirical Study of Horizon Length
Longer action horizons bottleneck LLM agent training through instability, but training with reduced horizons stabilizes learning and enables better generalization to longer horizons.
-
GUI Agents with Reinforcement Learning: Toward Digital Inhabitants
The paper delivers the first comprehensive overview of RL for GUI agents, organizing methods into offline, online, and hybrid strategies while analyzing trends in rewards, efficiency, and deliberation to outline a future roadmap.
-
AlphaEval: Evaluating Agents in Production
AlphaEval is a benchmark of 94 production-sourced tasks from seven companies for evaluating full AI agent products across six domains using multiple judgment methods, plus a framework to build similar benchmarks.
-
A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence
The paper delivers the first systematic review of self-evolving agents, structured around what components evolve, when adaptation occurs, and how it is implemented.
-
Beyond Static Evaluation: Building Simulation Environments for Scalable Agentic Reinforcement Learning
An architecture description of an enterprise RL gym for LLM agents with multi-dimensional reward shaping, presented through a customer support case study that contains no quantitative evaluation.
-
Large Language Model Agent: A Survey on Methodology, Applications and Challenges
A survey that deconstructs LLM agent systems via a methodology-centered taxonomy linking design principles to emergent behaviors, applications, and challenges.