pith. sign in

arxiv: 2606.05342 · v2 · pith:XMEP6RVRnew · submitted 2026-06-03 · 💻 cs.AI

SentinelBench: A Benchmark for Long-Running Monitoring Agents

Matheus Kunzler Maldaner , Adam Fourney , Amanda Swearngin , Hussein Mozannar , Gagan Bansal , Maya Murad , Rafah Hosn , Saleema Amershi This is my paper

Pith reviewed 2026-06-28 06:07 UTC · model grok-4.3

classification 💻 cs.AI
keywords AI agentsbenchmarkmonitoring taskslong-running tasksweb environmentsreaction timeresource usesustained attention
0
0 comments X

The pith

SentinelBench is a benchmark of 100 tasks that tests whether AI agents can monitor changing web environments patiently instead of acting continuously.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces SentinelBench to evaluate agents on long-running tasks that reward sustained attention over minutes or hours rather than constant tool use. It creates ten synthetic web environments that replay scripted event sequences, forcing agents to notice external changes and respond without wasting resources while idle. The benchmark records task completion, reaction time, and resource consumption to expose design tradeoffs. Baseline runs with three models and two browser harnesses show that agent choices produce large differences in these metrics.

Core claim

SentinelBench contains 100 tasks across ten synthetic web environments that expose live interfaces and replay scripted event sequences, requiring agents to navigate pages whose state changes over time. It measures task completion, reaction time, and resource use to quantify the tradeoff between responsiveness and cost, and baseline results across three models and two harnesses establish that the benchmark distinguishes meaningful differences in agent behavior.

What carries the argument

Time-evolving synthetic web environments with scripted event sequences that test sustained attention by requiring agents to wait for external changes before acting.

If this is right

  • Agent designs that favor waiting will record lower resource use while preserving high task completion rates.
  • Reaction time becomes a primary evaluation metric alongside accuracy for monitoring tasks.
  • Browser harness choices can produce large swings in measured cost and responsiveness.
  • Future agents can be ranked against the reported baselines on the same 100 tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar scripted-event environments could be built for non-web domains such as sensor streams or file-system monitoring.
  • The measured cost-reaction tradeoff may encourage hybrid agent loops that switch between idle and active modes based on predicted event timing.
  • Public leaderboards built on these tasks would make sustained-attention performance a visible optimization target for model developers.

Load-bearing premise

The ten synthetic web environments and their scripted sequences are representative enough of real-world long-running monitoring tasks to support claims about agent design tradeoffs.

What would settle it

Running the same agents on actual deployed monitoring tasks outside the ten synthetic environments and finding that relative performance rankings reverse or that resource and reaction metrics no longer correlate with task success.

Figures

Figures reproduced from arXiv: 2606.05342 by Adam Fourney, Amanda Swearngin, Gagan Bansal, Hussein Mozannar, Matheus Kunzler Maldaner, Maya Murad, Rafah Hosn, Saleema Amershi.

Figure 1
Figure 1. Figure 1: Agent execution timelines in SentinelBench for a representative scenario. Each row visualizes one agent strategy over simulated time as events are played back in the simulation (screenshots show the state of the app after selected events). The sleep agent spends most of its time executing fixed interval polling while the wait_for agent waits for a condition for an extended period and resumes once the envir… view at source ↗
Figure 2
Figure 2. Figure 2: Four SentinelBench environments. Each consists of a synthetic environment with a live web interface and a set of monitoring tasks with scripted sequences of events that can be played back to simulate the environment evolving over time. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Anatomy of a SentinelBench monitoring task: prompt, environment UI at the trigger moment, event timeline, and the underlying scenario JSON that drives the harness (from the MicroFy environment). serving a single client who is already authenticated. In other words, while they present a convincing facade, they should never be deployed as, or mistaken for, production web applications. Environment UI Surfaces … view at source ↗
Figure 4
Figure 4. Figure 4: SentinelBench includes 100 synthetic user personas. Each persona consists of core demographic and biographical attributes plus per-environment subprofiles that ground the same identity coherently across all 10 simulated applications (MicroGram, MicroScholar, MicroFy, MicroHub, MicroDin shown). Chris Taylor, shown here, is the principal user. I.e., the user whose profile is accessed by the agent. Taylor, th… view at source ↗
Figure 5
Figure 5. Figure 5: The simulation life cycle for SentinelBench environments. The evaluation harness interacts with the server to transition between life cycle states. For example, one task asked the agent to alert the user if any instant messages mentioned failures with a site’s payment-processing system. However, the scenario also included messages discussing problems on the site’s checkout page. These messages were meant t… view at source ↗
Figure 6
Figure 6. Figure 6: Per-task API cost in USD for each model and tool configuration. Box outlines show the interquartile [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-task API cost as a function of target event time for GPT-5.4 under the [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-task completion time as a function of target event time for GPT-5.4 under the [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The media-generation pipeline. An environment-independent core of LLM-generated personas [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MicroMail, modeled on Gmail/Outlook. Left: the populated inbox with an email open in the [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MicroChat, modeled on Slack/Teams. Left: the conversation list, with collapsible favorites and [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: MicroDin, modeled on LinkedIn. Left: the home feed. Right: the My Network page, with [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: MicroFy, modeled on Spotify. Left: the music library track grid. Right: an artist/playlist page. [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: MicroGram, modeled on Instagram. Left: a post in the feed. Right: the photo grid on a profile. [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: MicroHood, modeled on Robinhood. Left: the portfolio dashboard with its value chart. Right: [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: MicroHub, modeled on GitHub. Left: the repository overview. Right: the issues list. Catalog: [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: MicroLendar, modeled on Google Calendar. Left: the month grid. Right: the schedule/agenda [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: MicroScholar, modeled on Google Scholar. Left: a search results list. Right: an author profile [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: MicroTube, modeled on YouTube. Left: the subscriptions feed. Right: the video player with [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
read the original abstract

AI agents are increasingly asked to carry out work that spans minutes, hours, or longer. Yet the default model of agent behavior is continuous action: issuing tool calls, refreshing pages, searching for alternatives, or otherwise trying to force progress. This is the wrong approach for many long-running tasks, which are better served by a strategy of sustained attention. Instead, agents should monitor an environment, notice when an external event makes progress possible, then respond promptly without wasting resources while waiting. To measure progress on this class of tasks, we introduce SentinelBench, an open-source benchmark for time-evolving monitoring tasks. SentinelBench contains 100 tasks across 10 synthetic web environments, including email, calendars, finance, professional networking, and entertainment. Each environment exposes a live web interface and replays a scripted sequence of events, requiring agents to navigate and reason about web pages whose state shifts underfoot. SentinelBench measures task completion, reaction time, and resource use, exposing the tradeoff between responsiveness and cost. We report results across three models and two browser-agent harnesses, establishing performance baselines for future comparison and demonstrating how agent design choices can dramatically impact key metrics. Together, these results show that SentinelBench distinguishes meaningful differences in agent behavior.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces SentinelBench, an open-source benchmark with 100 tasks across 10 synthetic web environments (email, calendars, finance, etc.) that replay scripted event sequences. Agents must navigate live web interfaces, monitor for state changes, and respond promptly without continuous action. It evaluates three models and two browser-agent harnesses on task completion, reaction time, and resource use, establishing baselines and claiming that the benchmark distinguishes meaningful differences in agent behavior for long-running monitoring tasks.

Significance. If the benchmark's validity holds, it fills a gap by targeting sustained-attention strategies over continuous-action defaults, providing reproducible baselines and exposing responsiveness-cost tradeoffs. The open-source release and focus on time-evolving web tasks could support systematic progress on long-horizon agent design.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'these results show that SentinelBench distinguishes meaningful differences in agent behavior' rests on performance gaps across models/harnesses, yet the scripted event sequences (explicitly described as 'replays a scripted sequence of events') create a risk that agents optimize for detectable patterns or fixed timings rather than genuine responsiveness to unpredictable external events; this directly threatens whether the reported metrics test the intended monitoring tradeoff.
  2. [Abstract] Abstract and benchmark construction: the 10 synthetic environments are asserted to be representative of real-world long-running monitoring, but no justification, diversity metrics, or comparison to live unpredictable inputs is provided; without this, distinctions in completion/reaction/resource metrics may not generalize beyond the specific scripted replays.
minor comments (2)
  1. [Abstract] Abstract: 'establishing performance baselines for future comparison' would be strengthened by explicit mention of the exact models, harnesses, and aggregate metric values (even if detailed tables appear later).
  2. The description of environments as 'live web interface' while also 'replays a scripted sequence' creates minor ambiguity about whether the interface is truly dynamic or deterministic; clarify in the methods section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on the abstract and benchmark design. We address each point below, agreeing where the manuscript requires clarification or expansion and proposing targeted revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'these results show that SentinelBench distinguishes meaningful differences in agent behavior' rests on performance gaps across models/harnesses, yet the scripted event sequences (explicitly described as 'replays a scripted sequence of events') create a risk that agents optimize for detectable patterns or fixed timings rather than genuine responsiveness to unpredictable external events; this directly threatens whether the reported metrics test the intended monitoring tradeoff.

    Authors: We acknowledge the valid concern that scripted replays could allow pattern exploitation rather than true event-driven monitoring. However, the benchmark still requires agents to observe live web interfaces and react to state changes whose exact timing is not known in advance, which directly tests the distinction between sustained attention and continuous-action strategies. The reported gaps in reaction time and resource use reflect meaningful behavioral differences under these conditions. That said, we agree the setup does not fully address stochastic real-world events. We will revise the abstract to temper the claim and add an explicit limitations paragraph noting the controlled nature of the scripts and the value of future live-input extensions. This constitutes a partial revision. revision: partial

  2. Referee: [Abstract] Abstract and benchmark construction: the 10 synthetic environments are asserted to be representative of real-world long-running monitoring, but no justification, diversity metrics, or comparison to live unpredictable inputs is provided; without this, distinctions in completion/reaction/resource metrics may not generalize beyond the specific scripted replays.

    Authors: The environments were chosen to span common categories of time-evolving web tasks (email, calendar, finance, etc.), but we agree that the manuscript provides insufficient explicit justification, diversity metrics, or discussion of generalization. In the revised version we will expand the benchmark construction section with: (1) selection rationale tied to prevalence in real-world monitoring scenarios, (2) quantitative diversity measures such as event frequency variance and state-change counts per environment, and (3) a clearer statement that the synthetic replays provide reproducible baselines while not substituting for fully live, unpredictable inputs. This will be addressed with a full revision on this point. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark definition with direct empirical reporting

full rationale

The paper introduces SentinelBench as a new benchmark consisting of 100 tasks in 10 synthetic web environments with scripted event sequences, then reports measured task completion, reaction time, and resource metrics across three models and two harnesses. No equations, parameter fitting, derivations, or self-citation chains appear in the provided text. The central claim that the benchmark distinguishes meaningful differences rests on the explicit construction and execution of those tasks rather than any reduction to prior fitted inputs or self-referential definitions. This is a standard benchmark paper with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central contribution is the definition of the benchmark itself rather than any derivation from prior equations. No free parameters, axioms, or invented physical entities are introduced.

pith-pipeline@v0.9.1-grok · 5775 in / 1070 out tokens · 21980 ms · 2026-06-28T06:07:01.816322+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 15 canonical work pages · 7 internal anchors

  1. [1]

    arXiv preprint arXiv:2509.17158 , year=

    ARE: Scaling up agent environments and evaluations , author=. arXiv preprint arXiv:2509.17158 , year=

    work page arXiv

  2. [2]

    The Fourteenth International Conference on Learning Representations (ICLR) , url=

    GAIA2: Benchmarking llm agents on dynamic and asynchronous environments , author=. The Fourteenth International Conference on Learning Representations (ICLR) , url=

  3. [3]

    Proactive agent research environment: Simulating active users to evaluate proactive assistants.arXiv preprint arXiv:2604.00842,

    Proactive agent research environment: Simulating active users to evaluate proactive assistants , author=. arXiv preprint arXiv:2604.00842 , year=

    work page arXiv

  4. [4]

    and Barnes, Elizabeth and Chan, Lawrence , journal =

    Kwa, Thomas and West, Ben and Becker, Joel and Deng, Amy and Garcia, Katharyn and Hasin, Max and Jawhar, Sami and Kinniment, Megan and Rush, Nate and Von Arx, Sydney and Bloom, Ryan and Broadley, Thomas and Du, Haoxing and Goodrich, Brian and Jurkovic, Nikola and Miles, Luke Harold and Nix, Seraphina and Lin, Tao and Parikh, Neev and Rein, David and Sato,...

    2025

  5. [5]

    Is there a half-life for the success rates of

    Ord, Toby , year =. Is there a half-life for the success rates of. 2505.05115 , archivePrefix =

    work page arXiv

  6. [6]

    The Long-Horizon Task Mirage? Diagnosing Where and Why Agentic Systems Break

    The Long-Horizon Task Mirage? Diagnosing Where and Why Agentic Systems Break , author=. arXiv preprint arXiv:2604.11978 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning

    LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning , author=. arXiv preprint arXiv:2604.14140 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2604.00010 , year=

    Can LLMs Perceive Time? An Empirical Investigation , author=. arXiv preprint arXiv:2604.00010 , year=

    work page arXiv

  9. [9]

    Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks

    Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks , author=. arXiv preprint arXiv:2604.24964 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    2025 , month = jul, url =

    How Does Time Horizon Vary Across Domains? , author =. 2025 , month = jul, url =

    2025

  11. [11]

    2025 , url =

    Wijk, Hjalmar and Lin, Tao and Becker, Joel and Jawhar, Sami and Parikh, Neev and Broadley, Thomas and Chan, Lawrence and Chen, Michael and Clymer, Josh and Dhyani, Jai and Ericheva, Elena and Garcia, Katharyn and Goodrich, Brian and Jurkovic, Nikola and Karnofsky, Holden and Kinniment, Megan and Lajko, Aron and Nix, Seraphina and Sato, Lucas and Saunders...

    2025

  12. [12]

    2025 , url =

    Rein, David and Becker, Joel and Deng, Amy and Nix, Seraphina and Canal, Chris and O'Connel, Daniel and Arnott, Pip and Bloom, Ryan and Broadley, Thomas and Garcia, Katharyn and Goodrich, Brian and Hasin, Max and Jawhar, Sami and Kinniment, Megan and Kwa, Thomas and Lajko, Aron and Rush, Nate and Sato, Lucas Jun Koba and Von Arx, Sydney and West, Ben and ...

    2025

  13. [13]

    The Illusion of Diminishing Returns: Measuring Long Horizon Execution in

    Sinha, Akshit and Arun, Arvindh and Goel, Shashwat and Staab, Steffen and Geiping, Jonas , booktitle =. The Illusion of Diminishing Returns: Measuring Long Horizon Execution in. 2026 , url =

    2026

  14. [14]

    and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , journal =

    Zhou, Shuyan and Xu, Frank F. and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , journal =. 2023 , url =

    2023

  15. [15]

    2024 , url =

    Koh, Jing Yu and Lo, Robert and Jang, Lawrence and Duvvur, Vikram and Lim, Ming Chong and Huang, Po-Yu and Neubig, Graham and Zhou, Shuyan and Salakhutdinov, Ruslan and Fried, Daniel , journal =. 2024 , url =

    2024

  16. [16]

    2024 , url =

    He, Hongliang and Yao, Wenlin and Ma, Kaixin and Yu, Wenhao and Dai, Yong and Zhang, Hongming and Lan, Zhenzhong and Yu, Dong , journal =. 2024 , url =

    2024

  17. [17]

    2023 , url =

    Deng, Xiang and Gu, Yu and Zheng, Boyuan and Chen, Shijie and Stevens, Sam and Wang, Boshi and Sun, Huan and Su, Yu , booktitle =. 2023 , url =

    2023

  18. [18]

    2024 , url =

    Yoran, Ori and Amouyal, Samuel Joseph and Malaviya, Chaitanya and Bogin, Ben and Press, Ofir and Berant, Jonathan , journal =. 2024 , url =

    2024

  19. [19]

    Le Sellier De Chezelles, Thibault and Gasse, Maxime and Drouin, Alexandre and Caccia, Massimo and Boisvert, L. The. arXiv preprint arXiv:2412.05467 , year =

    work page arXiv

  20. [20]

    and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre , booktitle =

    Drouin, Alexandre and Gasse, Maxime and Caccia, Massimo and Laradji, Issam H. and Del Verme, Manuel and Marty, Tom and Vazquez, David and Chapados, Nicolas and Lacoste, Alexandre , booktitle =. 2024 , editor =

    2024

  21. [21]

    2024 , url =

    Xie, Tianbao and Zhang, Danyang and Chen, Jixuan and Li, Xiaochuan and Zhao, Siheng and Cao, Ruisheng and Hua, Toh Jing and Cheng, Zhoujun and Shin, Dongchan and Lei, Fangyu and Liu, Yitao and Xu, Yiheng and Zhou, Shuyan and Savarese, Silvio and Xiong, Caiming and Zhong, Victor and Yu, Tao , journal =. 2024 , url =

    2024

  22. [22]

    2023 , url =

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and Zhang, Shudan and Deng, Xiang and Zeng, Aohan and Du, Zhengxiao and Zhang, Chenhui and Shen, Sheng and Zhang, Tianjun and Su, Yu and Sun, Huan and Huang, Minlie and Dong, Yuxiao and Tang, Jie , journal =...

    2023

  23. [23]

    and Kang, Jikun and Wu, Wenqi and Christianos, Filippos and Greenlee, Fraser and Toulis, Andy and Purtorab, Marvin , journal =

    Thomas, George and Chan, Alex J. and Kang, Jikun and Wu, Wenqi and Christianos, Filippos and Greenlee, Fraser and Toulis, Andy and Purtorab, Marvin , journal =. 2025 , url =

    2025

  24. [24]

    ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

    ClawsBench: Evaluating capability and safety of llm productivity agents in simulated workspaces , author=. arXiv preprint arXiv:2604.05172 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2603.01357 , year=

    ASTRA-bench: Evaluating Tool-Use Agent Reasoning and Action Planning with Personal User Context , author=. arXiv preprint arXiv:2603.01357 , year=

    work page arXiv

  26. [26]

    The Agent's First Day: Benchmarking Learning, Exploration, and Scheduling in the Workplace Scenarios

    The Agent's First Day: Benchmarking Learning, Exploration, and Scheduling in the Workplace Scenarios , author=. arXiv preprint arXiv:2601.08173 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2511.01527 , year=

    TPS-Bench: Evaluating AI Agents' Tool Planning & Scheduling Abilities in Compounding Tasks , author=. arXiv preprint arXiv:2511.01527 , year=

    work page arXiv

  28. [28]

    AlphaEval: Evaluating Agents in Production

    AlphaEval: Evaluating Agents in Production , author=. arXiv preprint arXiv:2604.12162 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    2026 , url =

    Levy, Ido and Wiesel, Ben and Marreed, Sami and Oved, Alon and Yaeli, Avi and Shlomov, Segev , booktitle =. 2026 , url =

    2026

  30. [30]

    Conference on Language Modeling (COLM) , year =

    An Illusion of Progress? Assessing the Current State of Web Agents , author =. Conference on Language Modeling (COLM) , year =

  31. [31]

    International Conference on Learning Representations (ICLR) , year =

    Mialon, Gr. International Conference on Learning Representations (ICLR) , year =

  32. [32]

    2024 , url =

    Trivedi, Harsh and Khot, Tushar and Hartmann, Mareike and Manku, Ruskin and Dong, Vinty and Li, Edward and Gupta, Shashank and Sabharwal, Ashish and Balasubramanian, Niranjan , journal =. 2024 , url =

    2024

  33. [33]

    and Song, Yufan and Li, Boxuan and Tang, Yuxuan and Jain, Kritanjali and Bao, Mengxue and Wang, Zora Z

    Xu, Frank F. and Song, Yufan and Li, Boxuan and Tang, Yuxuan and Jain, Kritanjali and Bao, Mengxue and Wang, Zora Z. and Zhou, Xuhui and Guo, Zhitong and Cao, Murong and Yang, Mingyang and Lu, Hao Yang and Martin, Amaad and Su, Zhe and Maben, Leander and Mehta, Raj and Chi, Wayne and Jang, Lawrence and Xie, Yiqing and Zhou, Shuyan and Neubig, Graham , boo...

    2024

  34. [34]

    arXiv preprint arXiv:2508.09124 , year =

    Wang, Weixuan and Han, Dongge and Madrigal Diaz, Daniel and Xu, Jin and R. arXiv preprint arXiv:2508.09124 , year =

    work page arXiv

  35. [35]

    2026 , url =

    Zhao, Yujie and Yuan, Boqin and Huang, Junbo and Yuan, Haocheng and Yu, Zhongming and Xu, Haozhou and Hu, Lanxiang and Shankarampeta, Abhilash and Huang, Zimeng and Ni, Wentao and Tian, Yuandong and Zhao, Jishen , journal =. 2026 , url =

    2026

  36. [36]

    Empowering long-running

    Thacker, Joseph , howpublished =. Empowering long-running. 2024 , month = may, url =

    2024

  37. [37]

    2025 , url =

    Tasks in. 2025 , url =

    2025

  38. [38]

    2025 , month = jul, url =

    Introducing. 2025 , month = jul, url =

    2025

  39. [39]

    2026 , month = apr, note =

    Schedule recurring tasks in. 2026 , month = apr, note =

    2026

  40. [40]

    2025 , url =

    Yutori Scouts:. 2025 , url =

    2025

  41. [41]

    2025 , url =

    Building the Proactive, Multi-Agent Architecture Powering Scouts , author =. 2025 , url =

    2025

  42. [42]

    2025 , url =

    Introducing ambient agents , author =. 2025 , url =

    2025

  43. [43]

    From agent loops to structured graphs: A scheduler-theoretic framework for

    Hu, Wei , journal =. From agent loops to structured graphs: A scheduler-theoretic framework for. 2026 , url =

    2026

  44. [44]

    and Wutschitz, Lukas and Chen, Yanzhi and Sim, Robert and Rajmohan, Saravan , journal =

    Kang, Minki and Chen, Wei-Ning and Han, Dongge and Inan, Huseyin A. and Wutschitz, Lukas and Chen, Yanzhi and Sim, Robert and Rajmohan, Saravan , journal =. 2025 , url =

    2025

  45. [45]

    2022 , url =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , journal =. 2022 , url =

    2022

  46. [46]

    2025 , month = sep, url =

    Effective Context Engineering for AI Agents , author =. 2025 , month = sep, url =

    2025

  47. [47]

    2025 , url =

    Mozannar, Hussein and Bansal, Gagan and Tan, Cheng and Fourney, Adam and Dibia, Victor and Chen, Jingya and Gerrits, Jack and Payne, Tyler and Maldaner, Matheus Kunzler and Grunde-McLaughlin, Madeleine and Zhu, Eric and Bassman, Griffin and Alber, Jacob and Chang, Peter and Loynd, Ricky and Niedtner, Friederike and Kamar, Ece and Murad, Maya and Hosn, Raf...

    2025

  48. [48]

    2025 , month = oct, url =

    Tell me when: Building agents that can wait, monitor, and act , author =. 2025 , month = oct, url =

    2025

  49. [49]

    2025 , type =

    Sabotage risk pilot report , author =. 2025 , type =

    2025

  50. [50]

    arXiv preprint arXiv:2503.20314 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    2025 , url =

    Gong, Junmin and Zhao, Wenxiao and Wang, Sen and Xu, Shengyuan and Guo, Jing , howpublished =. 2025 , url =

    2025

  52. [52]

    2025 , month = nov, url =

    2025