Measuring human-agent teams in deployment

Agentic AI has moved from research demos to broad deployment. Coding agents handle substantial portions of professional software workflows (Anthropic, 2026), clinical decision-support agents triage patients (Taylor et al., 2025), and conversational agents mediate human learning and relationships (Shi et al., 2025; Zhang et al., 2025). The community's evaluation methods, built largely for static benchmark performance, have not kept pace.

Recent work finds that 83% of agentic-AI evaluations are dominated by technical metrics, while human-centered, safety, and economic dimensions remain peripheral (Jafari Meimandi et al., 2025). HAIC 2026 builds a methodological foundation for the empirical evaluation of human-agent teams. The central question is how to evaluate and govern human-agent systems rigorously as they coevolve with the people who use them, both in general deployment and in high-stakes domains such as healthcare, mental health, aviation, and finance.

This is the second edition of the HAIC series, following the inaugural ICLR 2025 workshop. Where the first workshop mapped coevolution broadly across five themes, this one commits to one focused operationalization, with each theme anchored in recent peer-reviewed evidence.

Submission instructions

The paper submission deadline is August 29, 2026, 11:59pm Anywhere on Earth. Papers are submitted through the HAIC 2026 OpenReview portal. Supplementary material is due at the same time as the main paper.

Papers may be revised as many times as needed up to the submission deadline. Revisions are not permitted while a paper is under review.

Paper length

The main text must take up at most 9 pages. This limit is enforced strictly: a submission with main text on the 10th page will be desk rejected. The limit applies to the initial submission and to the final camera-ready version alike.

References do not count toward the limit, and there is no cap on pages used for the bibliography. Authors may add as many pages of appendices after the bibliography as they wish, but reviewers are not required to read the appendix. The full nine pages are worth using only where larger figures or additional detail genuinely call for them.

Style files and templates

Submissions must use the NeurIPS 2026 LaTeX style files: Formatting Instructions for NeurIPS 2026 (ZIP, 20 KB). The archive contains neurips_2026.sty, the template neurips_2026.tex, and checklist.tex.

This workshop reviews double-blind, so load the style file with the dblblindworkshop option and nothing else. Omitting final and preprint is what anonymizes the paper and adds the line numbers reviewers refer to:

\usepackage[dblblindworkshop]{neurips_2026}
\title{Your paper title}
\workshoptitle{Human-AI Coevolution: Measuring Human-Agent Teams in the Agentic Era}

Workshop papers need \workshoptitle{} in addition to \title{}; it sets the footnote identifying the workshop. Accepted papers add final for the camera-ready version, \usepackage[dblblindworkshop, final]{neurips_2026}, which reveals author names and removes the line numbers. Do not use sglblindworkshop, which de-anonymizes the paper, and do not alter the margins, font sizes, or line spacing the style file sets.

Submissions must be anonymized. Author names, affiliations and acknowledgements are omitted, and references to the authors' own prior work are written in the third person.

Double-blind reviewing

Reviewing is double blind. Reviewers cannot see author names while conducting reviews, and authors cannot see reviewer names. Reviewer comments remain anonymous to authors.

Reviews are visible to the paper's authors and to the organizing committee, and are not published. Rejected submissions are not made public and are never deanonymized. Accepted contributions are posted publicly on OpenReview after the notification, and the program includes spotlight talks and posters of accepted papers.

Non-archival status and dual submission

The workshop is non-archival. Acceptance is not a publication of record, so authors may submit the same work to a conference or journal afterwards without prejudice.

Posting a submission to a preprint server such as arXiv is permitted, both before and during the review period. Work that has appeared on a non-peer-reviewed website, or that has been presented at a venue without published proceedings, may be submitted. Papers citing the authors' own earlier related work may be submitted, subject to the anonymization requirement above.

Use of large language models

Large language models may be used as a general-purpose assistive tool. Authors and reviewers take full responsibility for everything submitted under their names, including model-generated content that could be construed as plagiarism or scientific misconduct, such as fabricated results or references. Language models are not eligible for authorship.

Withdrawal

Authors may withdraw a submission at any time before the notification. Withdrawn submissions are removed from consideration and are not made public.

Code of conduct and ethics

All participants, authors included, are required to follow the NeurIPS Code of Conduct. Authors should also read the NeurIPS Ethics Guidelines, which apply to submission, reviewing and discussion alike.

A workshop code of conduct, setting out expectations for a respectful and inclusive environment, will be published here before the workshop.

Each submission receives three double-blind reviews via OpenReview. Reviews follow NeurIPS conflict-of-interest policy strictly. Organizers will not submit contributions to the workshop, nor will their immediate students or postdocs, and reviewers will not review papers from their own institutions or from organizations defined as conflicting under the NeurIPS policy.

The program committee comprises approximately 30 researchers, with domain coverage targeting one third agentic-AI evaluation and machine-learning methodology, one third human-robot interaction and human-agent teaming, and one third applied researchers with direct deployment experience in healthcare, aviation, finance, or related high-stakes settings. Each reviewer is capped at three papers.

The full program committee will be announced here once recruitment closes.