Agentic OS Foundations & Abstractions
- Minimal, portable abstractions and theoretical principles for building and composing an agentic OS
- Architectures, training, adaptation, and alignment of foundation models as first-class system components
Overview
Why an OS layer for agentic AI?
Agentic AI systems increasingly persist state, plan over long horizons, coordinate multiple models and tools, and even self-evolve. Yet today each capability is rebuilt from scratch inside framework-specific silos. The field needs a new OS-like layer that provides common abstractions for memory, scheduling, routing, governance, and reproducibility.
Recognition of this requirement has led to a proliferation of promising but largely independent efforts, each tackling a different piece of the agentic systems challenge. Designing the needed OS layer requires genuine co-design across the ML and systems communities. The ML community must ask how models should be trained, structured, and exposed as system components, not merely as API endpoints. The systems community must ask what abstractions, memory hierarchies, scheduling policies, and execution substrates are required to make agentic behavior reliable and governable. These questions are fundamentally coupled: the right abstractions depend on model capabilities, and model capabilities depend on system support.
Another excellent workshop series in the systems research community is OS for Agents. Our AgenticOS workshop sits at the intersection of machine learning and systems, with a focus on co-designing the foundations across both communities.
Scaling beyond pilots
Enterprise agentic systems are scaling beyond pilots but lack robust governance and reproducibility, risking fragmented framework-specific conventions.
Protocols at critical mass
Standardization (MCP, A2A) has reached critical mass, enabling definition of system-layer abstractions before defaults solidify.
A fragmented field
Relevant research is fragmented across ML, systems, and applications, motivating a dedicated venue to unify vocabulary, benchmarks, and collaboration.
Scope
Topics & research questions
The workshop will explore the following foundational questions:
- Should foundation models be trained differently to serve as system components?
- What are the minimal abstractions for building agentic systems?
- How should agentic memory representations, policies, and cross-layer optimizations be designed?
- How can agentic workloads be optimized across system layers?
- How should models be routed and composed under uncertainty?
- How should long-horizon agentic systems be designed?
- How can constrained self-evolution preserve safety, reproducibility, and governance?
- How should agentic systems be evaluated?
Our goal is to establish the conceptual foundations of an operating-system layer for agentic AI by bringing together researchers in machine learning, systems, and scientific applications to identify common abstractions, shared benchmarks, and a long-term research agenda.
Participate
Call for papers
AgenticOS: Co-designing Systems and ML Foundations of an OS Layer for Agentic AI
- Workshop website
- agentic-fmos.github.io
- Location
- Sydney, Australia
- Date
- December 11, 2026
- Submission link
- OpenReview
Call for submissions
Foundation models are evolving from stateless inference services into persistent, tool-using computational entities. As agentic AI systems grow in scale and capability, they require a common operating layer for memory, execution, scheduling, governance, and resource management. This workshop brings together the ML and systems communities to co-design the principles, abstractions, and implementations of an operating system for agentic AI, recognizing that future model capabilities and future systems abstractions must evolve together. We invite contributions that address the fundamental research challenges in building an operating system for agentic AI, including, but not limited to:
Memory, State & Storage
- Agentic memory: representations, retrieval and update policies, consistency, and cross-layer optimization
- Storage systems and file systems for agent state, context, and long-lived artifacts
Resource Management & Execution
- Cross-layer optimization for agentic workloads (compute, memory, network, cost, latency)
- Scheduling and resource management for agentic workloads, including cluster/GPU autoscaling under multi-agent contention
- Model routing, multi-model composition, and context management under uncertainty
Long-Horizon & Self-Evolving Agents
- Long-horizon execution: planning, persistence, checkpointing, and recovery
- Continual adaptation and constrained self-evolution
- System support for autonomous adaptation while preserving reproducibility, safety, and governance
Trust, Safety & Governance
- Trust, safety, and security for agentic systems: isolation, access control, and threat models specific to autonomous, tool-using agents
- Observability, provenance, and auditability of autonomous agent behavior
Evaluation & Deployment
- Evaluation methodologies, benchmarks, simulators, and testbeds
- Domain-specific agentic system deployment challenges and solutions
Submission guidelines
We welcome submissions in two formats:
Extended Abstracts
Extended abstracts may present visionary ideas, AgenticOS abstractions, or position papers that can spark interesting discussions on the workshop theme.
Length: Up to 2 pages of technical content.
Regular Papers
Regular papers should present original research, systems experience, empirical analyses, or deployment studies relevant to the workshop themes.
Length: Up to 6 pages of technical content.
We particularly encourage submissions that span both machine learning and systems, as well as experience reports, deployment lessons, and negative results that expose important open research challenges.
Submissions should adhere to the NeurIPS formatting guidelines (download the NeurIPS 2026 template).
References and appendices are not subject to page limits. However, the main paper must be self-contained, and reviewers are not required to consult the appendix. All submissions must be anonymized for double-blind review. Authors should remove names, affiliations, acknowledgments, and other identifying information from their submissions.
Papers must be original, unpublished work. As this workshop is non-archival and has no formal proceedings, works under review or planned for submission to other venues may also be submitted, provided they do not violate the policies of those venues. Accepted submissions will have the option of being published on the workshop website.
Review criteria
Submissions will be evaluated based on technical novelty, interest to the community, lessons learned, and relevance to AgenticOS now or in the future.
Camera-ready instructions for accepted papers
Camera-ready versions of accepted papers may include one additional page of technical content beyond the submission limits: up to 7 pages for regular papers and up to 3 pages for extended abstracts.
References and appendices have no page limits and do not count toward these technical-content limits.
The deadline to submit the camera-ready paper is October 19, 2026 (Anywhere on Earth, AoE).
Submit your work
Submit your paper through the AgenticOS workshop portal on OpenReview. For any queries, contact: agenticos.workshop@gmail.com.
Timeline
Important dates
Aug 29, 2026
Paper submission deadline (OpenReview)
Sep 29, 2026
Final notifications
Oct 19, 2026
Camera-ready paper deadline
Dec 11, 2026
Workshop day, NeurIPS 2026, Sydney
All deadlines are anywhere on Earth (AoE) unless stated otherwise. Dates are subject to change.
Program
Speakers
Keynote Speaker
Ion Stoica
UC Berkeley
Invited Speakers
Azalia Mirhoseini
Stanford University
Hongru Wang
University of Edinburgh
Bo Li
University of Illinois Urbana-Champaign
Gosia Steinder
IBM Fellow
Yogesh Simmhan
IISc Bangalore
Deshraj Yadav
Mem0
Mohamed Wahib
RIKEN R-CCS
Manish Gupta
Google DeepMind
Joe Zhou
MultiFi.ai
December 11, 2026
Schedule
A full-day program organized to maximize discussion and community building.
Select Abstract on any talk to read its details.
Morning
8:30 – 8:45
Opening
Opening Remarks
Suparna Bhattacharya
8:45 – 9:30
Keynote
Title to be announced
Ion Stoica
9:30 – 10:00
Invited Talk
Theory of Agent: Teaching Language Agents to Think and Act
Hongru Wang
Abstract
The talk will discuss a unified perspective on agentic AI, focusing on the allocation of epistemic effort between internal reasoning and external acting, as well as implications for building autonomous and self-evolving agents.
10:00 – 10:30
Invited Talk
The Trust Layer for AgenticOS: Evaluation, Alignment and Runtime Protection
Bo Li
Abstract
Foundation models have garnered widespread attention due to their impressive performance across a range of applications. However, our understanding of the trustworthiness and risks of these models remains limited. The temptation to deploy proficient foundation models in sensitive domains like healthcare and finance, where errors carry significant consequences, underscores the need for rigorous safety evaluations, enhancement, and guarantees. Recognizing the urgent need for developing safe and beneficial AI, our recent research seeks to design a unified platform to evaluate the safety of LLMs and multimodal foundation models (MMFM) from both regulatory compliance risk assessment principle, and use-case-driven principle such as toxicity, stereotype bias, adversarial robustness, OOD robustness, ethics, privacy, and fairness. Based on our understanding of risks of FMs, we have been focused on enhancing their safety through knowledge integration and provide safety guardrail and certifications.
In this talk, I will first outline our foundational principles for safety evaluation, detail our red teaming tactics, and share insights gleaned from evaluating foundation models on our DecodingTrust, AIR, and MMDT platforms for LLMs and MMFM, including close-source, open-source, and compressed models. Further, I will delve into our work on enhancing model safety, such as hallucination mitigation. I will also explain how knowledge integration helps align models and prove that the RAG framework achieves provably lower conformal generation risks compared to vanilla LLMs. Finally, I will briefly discuss our efficient and resilient guardrail framework for risk mitigation in practice.
10:30 – 11:00
Morning Break
11:00 – 11:15
Invited Talk
Virtual
The Harness is the Bug: Rethinking Agent Execution from the Ground Up
Gosia Steinder
11:15 – 12:00
Spotlights
Spotlight Papers
Authors
About 4 oral presentations: 8 min talk + 2 min Q&A each
12:00 – 1:00
Lunch Break
Afternoon
1:00 – 2:10
Posters
Poster Session
Authors
2:10 – 4:50
Special Session
Co-designing ML & Systems Foundations of AgenticOS: Current Reality & the Way Forward
2:10 – 2:15
Session Introduction
Ian Foster
2:15 – 2:30
Mini Keynote 1
The Golden Age of Model-Software-Hardware Co-design
Azalia Mirhoseini
Abstract
Inference demand is projected to grow 10–100x per year, and at the scale AI is now trained and served, merely building more datacenter capacity is impractical. Meeting this demand will require new sources of performance. In this talk, I will argue that customization through co-design is the key to unlocking those gains, and that the biggest wins will come from full-stack optimization across model, software, and hardware.
2:30 – 2:45
Mini Keynote 2
Evaluating the Pareto Frontier of Agent Memory
Deshraj Yadav
Abstract
Agents increasingly depend on what they remember across months of work, but memory evaluation has stayed stuck in a question-answer format. A question is itself a retrieval cue: it tells the agent a fact is needed and often which one, so the evaluation begins only after the hardest step has already been done for the system. Real agents acting under an instruction get no such hint. Agent memory should instead be measured through the actions agents take, with two things reported alongside accuracy. Cost and latency, because memory is an optimization problem and a system can buy accuracy by re-reading everything with a frontier model at a price no deployment would pay. And per-test certification in both directions, solvable with the relevant history and unsolvable without it, because audits of widely used benchmarks have turned up wrong answer keys, mislabeled evidence, and judges that accept most deliberately incorrect answers.
Early results across built-in memory, dedicated memory plugins, and full context on an open source harness show two plugins reaching nearly identical accuracy while differing by about 70% in cost and close to 3x in latency, exactly the tradeoff that accuracy-only reporting hides.
2:45 – 3:00
Mini Keynote 3
Orchestrating Agentic Systems across Edge and Cloud
Yogesh Simmhan
3:00 – 3:30
Coffee Break
3:30 – 3:45
Mini Keynote 4
Enhancing the Efficiency of AI Models and Agents
Manish Gupta
3:45 – 4:00
Mini Keynote 5
From Tokens to Tasks: Efficient Execution of Agentic AI at Scale
Mohamed Wahib
Abstract
Agentic AI shifts the performance challenge from serving individual model requests to executing dynamic workflows that interleave inference, tool execution, API calls, and data access. Optimizing inference alone is therefore insufficient to minimize task completion time. This talk presents our ongoing investigation into efficient execution of agentic workloads on heterogeneous computing platforms. We examine how CPU–GPU interactions, memory and I/O demands, and contention across concurrent tasks shape end-to-end performance, and explore scheduling and resource allocation strategies to improve throughput and latency. Using supercomputing platforms as a testbed, we investigate how these challenges evolve at scale and discuss their implications for the execution and resource management foundations of an agentic operating system.
4:00 – 4:50
Panel
Panel Discussion
Host: Ian Foster
- Azalia Mirhoseini
- Deshraj Yadav
- Yogesh Simmhan
- Manish Gupta
- Mohamed Wahib
- Joe Zhou
4:50 – 5:00
Closing
Best Paper Award & Closing Remarks
Tarun Kumar
Team
Organizers
Reviewers
Program committee
Reviewing policies
NeurIPS reviewing policies. Our workshop follows the official NeurIPS reviewing policies and code of conduct. Please review and adhere to the NeurIPS reviewer guidelines and the NeurIPS Code of Ethics. These policies apply to all reviews submitted for the workshop.
LLM policy. Reviewers should not use large language models (LLMs) to generate their reviews, in whole or in part. This includes drafting the overall assessment, summarizing submissions, or producing detailed comments. Uploading submissions or any confidential material to LLM services is also prohibited, as it may compromise the confidentiality of the review process. The papers can best benefit from your own expertise and careful assessment, and we greatly appreciate the time and judgment you bring to the review process.
Reviewing guidelines
The AgenticOS workshop aims to bring together the machine learning and systems communities around a shared vision for OS-like abstractions for agentic AI. We encourage reviews that evaluate submissions through this interdisciplinary lens.
Extended Abstracts
Visionary ideas, proposed AgenticOS abstractions (see the call for papers), and position papers meant to spark discussion at the workshop.
- Technical novelty
- Does the idea offer a genuinely new abstraction, framing, or perspective on the OS layer for agentic AI, even if unproven?
- Interest to the community
- Would this spark productive debate or disagreement among ML and systems researchers in the room?
- Lessons learned
- If based on early experience or a small-scale system, are the takeaways clearly articulated?
- Relevance to AgenticOS
- Does it speak to abstractions, memory, scheduling, governance, or evaluation for agentic systems, now or as a forward-looking bet?
Do not require Full experimental evaluation, comprehensive related-work coverage, or proof of feasibility. A well-argued speculative position is acceptable.
Reject if The abstract is a thin summary of a paper already fully developed, is out of scope, or makes no claim a reader could productively engage with or challenge.
Regular Papers
Original research, systems experience, empirical analyses, or deployment studies. Hold these to a genuine research bar, but weigh systems/empirical contributions and deployment experience as highly as purely novel algorithmic results, consistent with the workshop’s co-design mission.
- Technical novelty
- New abstraction, mechanism, system, or empirical finding, or a novel combination/application of existing ones in the agentic OS context.
- Interest to the community
- Would ML and systems researchers both learn something from this? Does it help unify vocabulary or benchmarks across the two communities?
- Lessons learned
- Are conclusions grounded in evidence (experiments, deployment data, ablations, or failure analysis) rather than assertion?
- Relevance to AgenticOS
- Does the contribution map clearly onto one or more workshop topics (abstractions, memory/state, resource management, long-horizon/self-evolving agents, trust/safety/governance, evaluation)? See the call for papers.
- Evidence
- Claims should be supported by evidence appropriate to the paper’s type (benchmarks, case studies, ablations, or deployment metrics).
- Methodology
- Described with enough detail to assess soundness, even under the 6-page limit (references and appendix may carry supporting detail).
- Negative results
- Acceptable, and should be scored on the quality and honesty of the analysis, not on whether the outcome was positive.
Reject if The contribution is incremental with no meaningful lesson for the community, is out of scope for the workshop themes, or lacks any evidence for its central claims.
Final note. We encourage reviewers to provide constructive, actionable feedback that helps authors improve their work, regardless of the recommendation. Since this is a workshop, we particularly value reviews that identify promising ideas and highlight opportunities for discussion within the AgenticOS community. Thank you again for helping us make the workshop review process thoughtful, fair, and engaging. Please reach out to the organizing committee if you have any questions during the review period.
Area Chairs
- Cheng TanNortheastern University
- Cong XuHewlett Packard Enterprise (HPE Labs)
- Kyle ChardUniversity of Chicago
- Lianjie CaoHewlett Packard Enterprise (HPE Labs)
- Mohamed WahibRIKEN R-CCS
- Pekka EnbergTurso
- Robert UnderwoodArgonne National Laboratory
- Sourangshu BhattacharyaIIT Kharagpur
- Yogesh SimmhanIISc Bangalore
Program Committee Members
- Aishwarya KamalGoogle
- Ankur SharmaDatabricks
- Arian RajeCarnegie Mellon University
- Ashish DsaArbor
- Diana AlvaradoAWS
- Dong LiUC Merced
- Eiko YonekiUniversity of Cambridge
- Elron BandelIBM Research
- Feijie WuPurdue University
- Gaurav Kumar AmbastaAmazon
- Haiyan YinA*STAR Centre for Frontier AI Research, Singapore
- Haitian JiangNew York University
- Hanshi SunByteDance
- Hubertus FrankeIBM Research
- Hui GuanUniversity of Massachusetts Amherst
- In GimYale University
- James CaddenIBM
- Jianming TongGeorgia Tech
- Jiayi WangOak Ridge National Laboratory
- Jingtong HuUniversity of Pittsburgh
- Jinyang LiNew York University
- Karthik SureshAdobe
- Lara MauriUniversità degli Studi di Milano
- Laurent BindschaedlerMPI-SWS
- Lorenzo SaniUniversity of Cambridge
- Martin FoltinHewlett Packard Enterprise (HPE Labs)
- Minghong FangUniversity of Louisville
- Rafael Ferreira Da SilvaOak Ridge National Laboratory
- Raghav SharmaWorkday
- Rishi SharmaMicrosoft
- Samantika SuryHewlett Packard Enterprise
- Sanjay Kumar PatnalaScale AI
- Saurabh KalikarGoogle
- Shiqiang WangUniversity of Exeter
- Shuli JiangCarnegie Mellon University
- Siddhartha JainUT Austin
- Stefanos LaskaridisAmazon
- Sutanay ChoudhuryPacific Northwest National Lab
- Tong XieUniversity of New South Wales
- Vasileios TsouvalasEindhoven University of Technology
- Vineet GundechaHewlett Packard Enterprise
- Wenyue HuaMicrosoft
- Xiuyu LiStepFun
- Yae Jee ChoGoogle
- Ye TianUniversity of California San Diego
- Yi LiUniversity of Texas Dallas
- Yiwei YangUniversity of California, Santa Cruz
- Yokila AroraWalmart
- Yu Bo GaoUniversity of Toronto
- Yuke WangRice University
- Yun ShenHewlett Packard Enterprise (HPE Labs)
- Yusheng ZhengUniversity of California, Santa Cruz
- Zengqing WuOsaka University
- Zexi LiCUHK
- Zhanhong JiangIowa State University
- Zhenheng TangHKUST
- Zhixu DuDuke University
- Zihao YeUniversity of Washington