Amazon researcher building long-horizon LLM agent architectures and benchmarks for evolving agent harnesses
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
3
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Output cannot be scored: this snapshot lacks a reliable total-publication measure for this OpenAlex-resolved profile. The axis would normally track cumulative works, venue types, and yearly productivity; the partial year-by-year counts on file are not a sufficient basis.
Impact—
Impact cannot be scored: the snapshot offers only h-index 3 and i10-index 2, too thin to characterize influence. This axis would measure citations, uptake, and reuse of the researcher's work across the community.
Focus0/100
The focus score of 0 signals a currently diffuse footprint: field distribution splits evenly (0.5/0.5) between LLM agent evaluation/machine learning and AI agents (LLM systems), while topic counts scatter across reinforcement learning in robotics (9), planning (5), explainable AI (4), topic modeling (3), and NLP (3). Sustained output in one line would sharpen this axis.
Collaboration54/100
The collaboration score of 54 reflects a moderate, team-embedded pattern: the OpenAlex record lists 14 distinct coauthors, and twelve recur across projects, including Zixuan Ke, Vaidehi Patil, Haizhou Shi, Semih Yavuz, and Shafiq Joty. Both 2026 catalog papers place him as a middle author with no corresponding role, suggesting contribution within larger group efforts rather than small-team or solo leadership.
Activity57/100
Activity scores 57 on a record that is episodic rather than steady: OpenAlex counts 1-3 works per year across 2017-2023, then a cluster of six works in 2026, with six works within the recent three-year window. The 2026 concentration — including both Evolve-cataloged arXiv papers — marks a clear current uptick after earlier sparse years.
Reproducibility38/100
Reproducibility averages 38 across two Evolve-scored papers. The long-horizon-agent architecture paper is rated low readiness — its ten-day autonomous RL reproduction is a single campaign without baselines — while EvoHarnessBench is rated medium, with solid design and mechanistic analysis but heuristic skill/agent labels and thin statistics. Both carry enough released detail to be checked, though with caveats.
Research summary
Anurag Koul is a researcher at Amazon (United States) whose profile, name-resolved from OpenAlex bibliographic data, centers on large language model agents and how to evaluate them. The Evolve catalog includes two 2026 arXiv papers: 'An Architecture for Long-Horizon Agents: Levels, Ticks and Cascaded Intelligence,' which describes a levels/ticks/cascade harness that ran a ten-day autonomous reinforcement-learning reproduction with daily human checks, and 'EvoHarnessBench,' a first benchmark probing whether LLM agents keep pace with an evolving harness. On both he appears as a middle author without corresponding status. The bibliometric snapshot shows a compact record: h-index 3, i10-index 2, and works scattered across 2017-2023 before a cluster of six works in 2026, with six works in the recent three-year window. Topic frequencies place reinforcement learning in robotics first (9), followed by AI-based problem solving and planning (5) and explainable AI (4). Field distribution splits evenly between LLM agent evaluation/machine learning and AI agents (LLM systems). He has 14 distinct coauthors, including recurring collaborators such as Zixuan Ke, Vaidehi Patil, and Haizhou Shi. Reproducibility scoring across two papers averages 38, with the harness-architecture paper rated low readiness and EvoHarnessBench medium.
Collaboration pattern
The pattern is team-embedded rather than solo: 14 distinct coauthors appear in the OpenAlex record, and a recurring circle of twelve spans Amazon and academic collaborators — Zixuan Ke, Vaidehi Patil, Haizhou Shi, Yang Li, Ye Liu, Sarath Shekkizhar, Jiayu Wang, Xuan Phi Nguyen, Semih Yavuz, Mohit Bansal, Shafiq Joty, and Erik Nijkamp. On both 2026 catalog papers he sits as a middle author, not corresponding, consistent with a contributing specialist within larger LLM-agent teams.
Research keywords
Highlights
Levels/ticks/cascade harness runs a ten-day autonomous RL reproduction
EvoHarnessBench: first benchmark for agents facing an evolving harness
Timeline
- 2017Earliest indexed work appears in the OpenAlex record (1 work recorded for the year).
- 2018One indexed work, continuing an intermittent publication cadence.
- 2020Three indexed works — the highest single-year count in the early record.
- 2021-2023Sparse stretch with one indexed work in each of 2021, 2022, and 2023.
- 2026Six works appear, including both Evolve-cataloged arXiv papers: the long-horizon agent architecture and EvoHarnessBench.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric indicators and topics are name-resolved from OpenAlex only; no ORCID was available to verify identity merging for this name.
- No citation-impact dataset is present: the output and impact radar axes are unscored, and h-index (3) and i10-index (2) are the only influence signals on file.
- Only the two 2026 arXiv preprints have Evolve review verdicts and reproducibility scores; other indexed works were not assessed.
- Both cataloged papers are arXiv preprints with middle authorship; peer-review venue status is not captured in the data.