Managing Agents that Manage Agents

Workshop on Responsible Use of Meta-Agents

NeurIPS 2026 · December 11, 2026 · Sydney, Australia

About The Workshop

Most agentic systems today rely on manual design for their training, harnesses, and goals. As models grow more capable, new systems are emerging to automate this. AlphaEvolve writes and tests its own programs; GEPA and Meta-Harness use agents to optimize the trajectories of other agents, beating reinforcement learning at a fraction of the cost. We call these higher-order agents that build, optimize, and supervise other agents meta-agents.

Meta-agents will likely play an increasingly central role in how agentic systems are used and developed. Yet discourse on them stays dispersed across separate communities, from meta-learning to reinforcement learning to systems research. A systematic, interdisciplinary view is needed to address their potential for both use and misuse, convening core figures from these subfields alongside researchers in AI and organizational science.

The shift from manual to automated agent design raises open technical problems. A meta-agent can design the harness, train the agent system, and enable continual learning and self-improvement. When does a discovered or self-improved design actually generalize, and which optimization methods and learning signals improve a meta-agent's capabilities?

As capabilities scale, societal impact needs careful oversight. A meta-agent acts like a manager, decomposing objectives and assigning tasks to worker agents, so a misaligned one can spawn and coordinate swarms of subordinate agents toward the wrong goal. Meta-agents can also manage humans, with risks of economic disruption and eroded autonomy. So every result raises the same question: who evaluates and oversees an agent that builds or improves another agent, or itself? The workshop covers the full lifecycle, design, training, evaluation, deployment, and oversight, and every submission carries a responsible-use statement.



Topics

We invite work across the topics below, capability first, then evaluation, oversight, and governance. Topics include but are not limited to:

(1) Automated Design and Discovery of Agent Harnesses:

Most agent harnesses are still built by hand. We invite work on searching the space of agent architectures, tools, and harnesses (Meta Agent Search, automated harness design), and on telling when a discovered design generalizes past the tasks it was tuned on.

(2) Optimization and Post-Training of Compound Agentic Systems:

Once an agent is more than one model call, the problem is optimizing the whole program. We seek work on optimizers for multi-module systems (DSPy, GEPA, Trace), credit assignment across modules, post-training of compound systems, inference-time search, and context optimization.

(3) Self-Improvement and Open-Ended Evolution:

A meta-agent can point at itself. We welcome work on agents that rewrite their own code (the Darwin Gödel Machine, self-taught optimizers), self-play and co-evolution, open-endedness, and honest analyses of how self-improving agents drift or game their own reward.

(4) Evaluation and Benchmarks for Meta-Agents:

Does a discovered or self-improved agent actually generalize? We want benchmarks and tests for meta-level methods, the ones that answer this question rather than leaderboards that hide the failure cases.

(5) Misalignment and Safety for Meta-Agents:

Making agent execution observable and auditable, red-teaming systems that act on other agents, and detecting reward gaming, over-optimization, and drift before a system acts on other agents at scale.

(6) Governance and Human Oversight of Meta-Agents:

Rollback and halt controls a deployer can actually trigger, who signs off before a discovered agent ships, and how to govern agents that manage other agents, including perspectives from management science and other fields outside AI.



Call For Papers

Managing Agents that Manage Agents (NeurIPS 2026) invites submissions on architectures, algorithms, theory, empirical studies, benchmarks, and position papers about agents that design, optimize, supervise, train, or improve other agents. Submissions must present original, unpublished work that has not appeared at NeurIPS or other archival machine-learning venues.

Key Dates

  • Submission Deadline: August 29, 2026 5th Sept, AoE
  • Notification: on or before September 29, 2026, AoE
  • Camera-Ready Deadline: November 20, 2026, AoE
  • Workshop Date: December 11, 2026 (Sydney)
All deadlines follow the Anywhere on Earth (AoE) timezone.

Submission Site

Submissions are managed via OpenReview and remain private during review. All authors should maintain up-to-date OpenReview profiles for conflict-of-interest management and paper matching. Submit your paper at the OpenReview submission portal.

Interested in serving as a reviewer or Area Chair (AC)? Self-nominate using this form.

Camera-Ready Instructions

Camera-ready papers are due November 20, 2026, AoE. All accepted papers must use the Meta-Agents Workshop NeurIPS 2026 LaTeX template (ZIP). A compiled example PDF is also available.

The template is configured for camera-ready papers, with the Meta-Agents Workshop notice, author names and affiliations visible, and review line numbers removed. Keep the final option enabled, restore any author-identifying information and acknowledgments removed for review, and include the required responsible-use statement. The NeurIPS main-conference paper checklist is not required. Upload a single PDF, including references and any appendices, through OpenReview.

Full camera-ready papers may contain at most 10 pages of main text, including figures and tables; short camera-ready papers may contain at most 5 pages. References and appendices do not count toward these limits, but the main text must be self-contained. For initial submissions, the page limits below apply.

Scope

We welcome contributions across the topics above. Accepted papers are presented as posters, with a subset selected for oral or spotlight talks, and we give a Best Paper award and a Best Social Impact Paper award for the work that most thoughtfully addresses the societal implications of meta-agents. The workshop is in person at NeurIPS 2026 in Sydney.

Submission Guidelines

Formatting
Papers must be in English and use the NeurIPS 2026 LaTeX template customized for the Meta-Agents Workshop. The NeurIPS paper checklist is not required. Papers are submitted as a single PDF:
  • Full Papers: at most 9 pages (main text)
  • Short Papers: at most 4 pages (main text)
  • Demo Track: live demonstrations of meta-agent systems and tools, presented alongside the poster sessions.
  • Position Papers: on the governance, oversight, and societal impact of meta-agents.
  • References and appendices do not count toward the page limit, but the main text must be self-contained.
Responsible-Use Statement

Every submission also includes a short responsible-use statement covering the potential societal impacts of the proposed work and suggested mitigations. It is reviewed with the paper, and a missing one is grounds for desk rejection.

Review-Stage Anonymity
The workshop uses double-blind review for initial submissions, with author names, affiliations, and acknowledgments removed and prior work cited in the third person. Accepted camera-ready papers must include author names and affiliations, as described in the camera-ready instructions above.

Non-Archival Policy

The workshop is non-archival. Papers under review elsewhere are welcome, and accepted papers may be published at other venues afterward. Work already published at NeurIPS or other archival ML venues should not be submitted.

Contact
Email meta-agents-workshop@googlegroups.com.

Speakers & Panelists

Meet the invited speakers, panelists, and moderator.

Graham Neubig
Graham Neubig

Carnegie Mellon University

Speaker

Chelsea Finn
Chelsea Finn

Stanford University

Speaker

Jenny Zhang
Jenny Zhang

Recursive Superintelligence & UBC

Panelist

Bo Li
Bo Li

University of Illinois Urbana-Champaign

Speaker

Yuandong Tian
Yuandong Tian

Recursive

Speaker

Chen Sun
Chen Sun

Google DeepMind

Panelist

Ying Liu
Ying Liu

Scale AI

Speaker

Jess Bergs
Jess Bergs

UK AI Security Institute

Panelist

Rishi Bommasani
Rishi Bommasani

Stanford HAI

Panelist

Yi Tay
Yi Tay

Google DeepMind

Panelist

Christopher D. Manning
Christopher D. Manning

Stanford University

Moderator

Schedule

Friday, December 11, 2026 · Sydney time (AEDT). All invited talks include Q&A.

Time (AEDT) Session Participants
09:00 – 09:10Opening RemarksOrganizers
09:10 – 09:35Invited Talk 1Graham Neubig (Carnegie Mellon University)
09:35 – 10:20Oral Session 1Three selected authors
10:20 – 10:40Coffee Break
10:40 – 11:05Invited Talk 2Ying Liu (Scale AI)
11:05 – 11:30Invited Talk 3Chelsea Finn (Stanford University)
11:30 – 12:30Poster & Demo Session 1
12:00 – 13:00Attendee Lunch Window
13:00 – 13:25Invited Talk 4Bo Li (University of Illinois Urbana-Champaign)
13:25 – 14:10Oral Session 2Three selected authors
14:10 – 14:30Coffee Break
14:30 – 14:55Invited Talk 5Yuandong Tian (Recursive)
14:55 – 15:40Panel DiscussionGovernance of Meta-Agents
Jenny Zhang, Chen Sun, Jess Bergs, Rishi Bommasani, and Yi Tay
Moderator: Christopher D. Manning
15:40 – 16:40Poster & Demo Session 2
16:40 – 17:00Awards & Closing RemarksChristopher D. Manning & organizers

Organizers

This workshop is organized and advised by

Simon Yu
Simon Yu

Northeastern University

Dilara Soylu
Dilara Soylu

Stanford University

Derek Chong
Derek Chong

Stanford University

Ananjan Nandi
Ananjan Nandi

Stanford University

Apurva Gandhi
Apurva Gandhi

Carnegie Mellon University

Jiuding Sun
Jiuding Sun

Stanford University

Zichen Liu
Zichen Liu

Google DeepMind

Christopher D. Manning
Christopher D. Manning

Stanford University

Weiyan Shi
Weiyan Shi

Northeastern University

Sponsors

We gratefully acknowledge the support of our sponsors.

Scale AI

Scale AI

Meta

Meta

E2B

E2B