Overview
AI agents are rapidly evolving from experimental prototypes to always-on services that autonomously plan, invoke external tools, collaborate, and continuously interact with their environment. This shift challenges traditional operating system abstractions—processes, threads, files, sockets, and resource controllers—which were never designed for dynamic, semantically rich, adaptive agent workloads. To support AI agents at scale, operating systems themselves must become agentic, adapting their abstractions and resource management policies to the semantic behaviors of agents.
The second AgenticOS workshop, co-located with SOSP 2026, seeks original position papers and experience reports that explore OS-level mechanisms for AI-agent workloads. Our goal is to define the primitives, isolation models, scheduling techniques, and observability mechanisms necessary to build operating systems explicitly tailored to agent-based systems.
Workshop Schedule
September 29, 2026 — Afternoon session (local time)
Keynote Abstract & Speaker Bio
Abstract
Consider three failures. An agent places a supplier order through a web portal, the request times out after the order has already been accepted, and the agent retries. Two orders ship, and the task is scored a success. Another agent compacts its context, silently drops a standing rule, and deploys to the release branch without noticing. A human approves a pull request, a rebase changes the reviewed code, the workflow retries and reuses the original approval, and unreviewed code is merged. In each case the model's decision was reasonable and the outcome was wrong. Rules were broken, but no runtime contract stated them.
Agents add a problem that processes did not pose. An operating system never judged whether a program's actions made sense. It confined the program and left that judgment to whoever wrote it. An agent generates its own actions, and the guarantees the author used to carry are not in the model. That judgment now falls to the runtime between the model and the world, whether it is called a harness or an agent operating system. What that runtime must guarantee is the interface we lack, a kind of POSIX for agents.
The field already has many pieces of an agent operating system (sandboxes, schedulers, memory managers, kernels) but no shared account of what they must guarantee, and task success alone cannot show whether a guarantee held. In this talk, I will sketch that interface: four contracts, effect, state, composition, and intent, at different levels of maturity. For each I will show what we can state today, an early mechanism where we have one, and the evidence needed to check it.
Speaker Bio
Laurent Bindschaedler leads the Data Systems Group at the Max Planck Institute for Software Systems, where his work sits at the intersection of operating systems, databases, and machine learning. His group builds systems for massive-scale data management and AI, currently focused on agentic AI infrastructure, LLM-enhanced databases, and benchmarking methodologies for adaptive systems. He holds a PhD from EPFL and was a postdoctoral fellow at MIT CSAIL. His work has appeared at venues including SOSP, ASPLOS, EuroSys, ICML, EMNLP, and NDSS.
Topics of Interest
Topics of interest include (but are not limited to):
- New OS abstractions for agent execution (process/container/multikernel enhancements)
- Dynamic sandboxing and lightweight runtimes for securely executing agent-generated code and tasks
- Semantics-aware resource management and scheduling for dynamic, multi-agent workloads
- Long-lived state abstractions for managing agent context, prompts, and episodic memory
- Observability, provenance, and debugging for agent executions
- Compiler–OS co-design for adaptive and agent-aware JIT execution strategies
- GPU and accelerator virtualization for large-scale deployment of agent workloads
- Security and isolation mechanisms for agent-invoked tools, code, and data flows
- Agents managing systems: kernel tuning, anomaly detection, resource orchestration, failure recovery, and dynamic policy updates
- Inter-agent communication patterns and primitives
- Reliability and fault tolerance of agents
Submission Guidelines
We solicit two types of submissions:
Vision Papers
1–2 pages (excluding references).
Suitable for early-stage ideas, position statements, ongoing projects, demos, and insights from production systems. We strongly welcome contributions from industry practitioners and the open source community to share real-world experiences and challenges.
Research Papers
Up to 6 pages (excluding references).
Suitable for more complete concepts, research results and experience reports.
All submissions must follow the ACM double-column conference format. The review process is double-blind, with each submission receiving at least two reviews. Submit your paper via HotCRP.
Frequently Asked Questions
Can I submit work that is already on arXiv?
Yes. Posting on arXiv or other preprint servers does not disqualify a submission.
Can I submit work-in-progress or concurrent submissions?
Yes. We welcome early-stage ideas and ongoing work. Submissions under review at other venues are acceptable.
Will this affect future publication at other venues?
No. This workshop has no formal proceedings. Accepted papers will only appear on the workshop website, which does not preclude future publication at conferences or journals.
