Proposed workshop · under review

Agent Institutions

Designing institutional scaffolds for society-scale human–agent coordination

DAI 2026 · 8th Intl. Conference on Distributed AI City University of Hong Kong 29 November 2026

Overview

From the agent to the scaffold

The dominant research program in agentic AI is the improvement of the individual agent. But the scale now anticipated — billions of people interleaved with orders of magnitude more artificial agents — poses a coordination problem that no amount of individual capability resolves, and that legal and behavioural guardrails were never built to carry.

Durable coordination at society scale has never come from better individuals. It comes from structure: role-defined scaffolds that produce robust, repeatable decisions which other institutions can accept as legitimate inputs.

A courtroom, a peer review process, a parliamentary procedure — each reliably produces outcomes better than the people staffing it. The intelligence is invested in the scaffold rather than the occupant: in the separation of roles, the rules of evidence, the adversarial structure, the appeal path, the accumulated precedent. Markets are the one such structure our field has taken seriously, and their flatness and emergent order are genuinely powerful — but a purified economics cannot by itself build a society. Standing, membership, admissibility, the right to appeal, and the legitimacy of an outcome to those bound by it are not exchangeable value.

This workshop proposes that the design of agent institutions — nested, role-defined, populated by humans and agents alike — is the load-bearing problem for society-scale multi-agent AI, and that it is buildable and measurable now rather than a matter for speculation.

Three curves have crossed to make this urgent. Multi-agent deployments have moved from demonstration to production, and their characteristic failures are no longer reasoning failures but organizational ones — role confusion, unbounded delegation, responsibility diffusion, unauditable decision chains. An interoperability substrate has arrived fast enough that the question has shifted from whether agents can interoperate to under what procedures, with what standing, and answerable to whom. And the governance literature has converged on the finding that risks in agent collectives are structurally distinct from single-agent risks, and are not addressed by per-model alignment.

Each points at the same missing layer. That layer has a name in the older multi-agent systems literature — institution — which the current agentic wave has largely rediscovered without inheriting. This workshop convenes the multi-agent systems, negotiation, mechanism design, agent protocol, and governance communities around a single shift in the unit of analysis.

SCAFFOLD IMPARTIALREFEREE ADVOCATE“YES” ADVOCATE“NO” IMPARTIALJURY
An institutional scaffold. Roles are defined so as to facilitate cooperation, adversariality, and legitimate decisions. What matters is not the individual occupants but the intelligence built into the structure.

The organizing question

What transfers, and what breaks

Human institutions are the only large-scale, long-tested corpus of coordination designs we have. They are also built on assumptions about their participants that agents systematically violate. Sorting the transferable from the untransferable is the field's most useful near-term project, and it is this workshop's spine.

The Meaning Alignment Institute's AGI Institutions matrix gives the best available answer to the prior question — which institutions a world with powerful AI needs, arranged by scale from dyadic to global and by informational basis from protocols through rights, incentives, and norms to thick commitments. Read the three columns below as the next question put to every cell in that matrix: agent negotiation and bargaining, agent-to-agent mediation and contracts, autonomous corporate bylaws and grievance boards, deliberative value elicitation, AI commons management, international tribunals. For each, which part of the human original survives translation to agents, which part quietly stops working, and which part has no human precedent at all?

Transfers well

The portable patterns

Role separation. Prosecution and defence; author, reviewer, editor. Splitting a decision into roles with distinct information and incentives is substrate-independent.

Structured adversariality. Requiring a claim to survive a party motivated to destroy it is a general epistemic technology, not a human quirk.

Appeal and escalation. A procedure with no route to reconsideration is brittle regardless of who executes it.

Evidentiary standards. Rules about what may be considered and what must be disclosed — which agents can satisfy far more completely than humans, since traces can be recorded in full.

Registries and standing. Checking who is entitled to act in a role is easier with cryptographic identity than with human credentials.

Precedent. Prior decisions as a compressed store of institutional learning.

Breaks

The false friends

Deliberation-by-delay. Cooling-off periods, multiple readings, sleeping on it — the delay is doing epistemic work. Machine speed voids the mechanism while appearing to follow the procedure.

One member, one vote. Assumes members are costly to create. Agents can be forked, so quorums, juries, and majority voting degenerate.

Independence of judgment. An impartial jury assumes jurors fail independently. Twelve instances of one base model fail together, and the aggregation that makes juries work disappears.

Sanction and deterrence. Fines, disbarment, and shame all presuppose a persistent self that bears cost across time. An agent can be terminated and re-instantiated.

Oaths and internalized norms. A commitment made in one session may have no purchase in the next.

Consent and representation. Legitimacy derives from the consent of the governed — but agents are not principals, leaving the source of legitimacy for agent-made decisions unresolved.

No analogue

Native to agents

The harness as constitution. Tool access, context, session continuity, and memory permissions already determine what an agent may do and know. Nothing in human institutional theory corresponds to it.

Verifiable procedure. Execution traces mean compliance can be checked rather than trusted — dissolving a problem human institutions spend enormous resources on.

Counterfactual branching. A deliberation can be forked, run to completion along several branches, and compared. Human institutions get one run.

Institutions testable before deployment. A scaffold can be simulated against adversarial occupants before it governs anything real: constitutional design with an empirical loop.

Literal role swapping. The invariance property can actually be measured rather than asserted.

Cryptographic delegation. Authority scoped, attenuated, and revoked with a precision paper-based delegation never permitted.

We solicit work that adds to, contests, or empirically tests any row in these three lists — including negative results, where a human institutional form was tried on an agent population and failed, with a diagnosis of why.

Scope

Topics

We solicit contributions across seven areas, stated as problems a distributed AI audience can attack.

01

Scaffolding

Organizational and normative multi-agent systems revisited for the LLM era. Electronic institutions; role, norm, and protocol specification languages; institutional descriptions agents can read, negotiate, and be bound by. The harness as institution: tool access, session continuity, memory commons.

02

Negotiation, bargaining, contracts

Automated negotiation and bargaining protocols between agents; agent-to-agent contracting and enforcement; mediation and arbitration between agents acting for conflicting principals; commitment devices and escrow; renegotiation and breach. And the hard question underneath: what a contract even means between parties that can be duplicated or terminated at will.

03

Automated mechanism design

Mechanism design as a component of institutional design rather than a synonym for it. Automated and learning-based mechanism design; auction, matching, and market design where participants are agents; incentive compatibility under duplicable identities; algorithmic collusion among LLM agents. We especially want papers mapping the boundary where incentive alignment stops being sufficient and procedural or normative structure must take over.

04

The limits of flatness

Coordination over goods poorly represented as exchangeable value — standing, membership, admissibility of evidence, the right to appeal, the legitimacy of an outcome to those bound by it. Commons regimes and polycentric governance for compute, data, memory, and attention.

05

Generative adversariality

Structured conflict as a design material. Adversarial debate under an impartial referee; prosecution and defence as separated roles; red-teaming as a standing institutional function rather than a pre-deployment activity. And its failure modes: collusion, capture, escalation.

06

Role-based interchangeability

The measurement problem of the field. Institutional performance should be a property of the template, not the occupant. Metrics, benchmarks, and testbeds in which the object of evaluation is the scaffold — and results on when role interchangeability across humans and agents is achievable, and when it is not.

07

Nesting and legitimacy

Institutions whose reliable outputs are another's necessary inputs. Composition and interface standards; meta-institutions, appellate structures, registries. Accountability when consequential decisions are made by processes that are opaque, non-human, and too fast for meaningful oversight of individual outputs.

08

Cultural interoperability

Institutional pluralism as a technical design constraint. Conflicting legal ontologies and normative traditions as an interoperability problem. The risk of shallow skeuomorphism — bolting agents onto institutions designed for other eras of information exchange — and the reverse channel, where new hybrid forms reform the cultures they emerged from.

09

Institutions of knowledge

Peer review is the canonical institutional scaffold, and it is being rebuilt in public. AI-assisted and AI-participant review; credit and disclosure regimes; reproducibility infrastructure; the design of scientific institutions that admit agents as participants.

Benchmark

The Institution Design Challenge

Ahead of the workshop we release scenario environments in which individually strong agents reliably fail for structural rather than capability reasons — allocation of a depleting shared resource under private information; a dispute between principals whose agents hold conflicting instructions and share no escalation path; multi-round delegation where liability for an error must remain assignable after the fact.

Teams submit a scaffold: roles, procedures, admissible communications, decision rules, and appeal paths — as code plus a short institutional specification.

Invariance under occupant substitution

Entries are scored on outcome quality, but the criterion we want to introduce with this challenge is the role-occupancy swap test. Each scaffold is evaluated repeatedly with its role occupants swapped: across model families, across capability tiers, with a human in one role, and with a deliberately degraded or adversarial occupant in a randomly chosen role.

A good institution degrades gracefully. A scaffold whose measured performance is really the performance of its strongest occupant is, by this criterion, not an institution at all.

The harness, environments, and baseline scaffolds will be released publicly and maintained regardless of how many entries the challenge attracts.

Format

Program

A full day combining invited talks, contributed papers, a benchmark session, a live experiment, and community roadmapping. The schedule below is provisional and subject to the slot allocated by the DAI 2026 chairs.

TimeSession
09:00Framing: The Limits of Flatness — the case for the scaffold as unit of analysis
09:20Invited talk I — institutional and organizational multi-agent systems: what the field already solved, and what LLM agents broke
10:00Contributed session I — scaffolds, roles, and protocols
11:00Break
11:20Invited talk II — agent economies and the limits of market coordination
12:00Posters and demos — institution testbeds and environments
13:00Lunch
14:00Institution Design Challenge — results, baselines, and the role-occupancy swap analysis
15:00Live adversarial scaffold — one contested question put through referee, adversarial debate, and jury, with humans and deployed agents occupying the roles, and the composition varied mid-session
16:00Break
16:20Contributed session II — legitimacy, accountability, and cultural interoperability
17:00Roadmapping breakouts — drafting a public registry of institutional primitives
17:40Closing panel and synthesis

Outputs

  • Workshop report — the primitives registry, challenge results, and the transcript of the live scaffold session, released openly within eight weeks.
  • Challenge harness and environments — open source, released before the workshop and maintained after it.
  • Contributed papers — non-archival; authors retain full rights.

Submissions

Call for papers

The call opens if and when the workshop is accepted by the DAI 2026 chairs. Details below are the intended shape, published early so that prospective authors can plan.

  • Format — 4-page extended abstracts, plus unlimited references and appendices, in the DAI 2026 style.
  • Non-archival. Work may be concurrently under review elsewhere. This is deliberate: it raises the quality available to a first-edition workshop and lowers the barrier for legal, governance, and humanities scholars whose publication norms differ from those of machine learning.
  • Review — double-blind, two reviews per submission, organizers arbitrating. Reviewing norms, including disclosure rules for AI-assisted review, will be published alongside the call. Given the subject matter, we intend to treat our own review process as a documented instance of institutional design and to report on it.
  • Presentation — accepted work appears as short talks and in the poster and demo session.
  • Registration — all participants register for DAI 2026.

We particularly welcome submissions that would not find an easy home at a purely technical venue: institutional analysis, legal and normative design, and empirical studies of deployed agent organizations, alongside systems and benchmark work.

Schedule

Key dates

DateMilestone
30 Jul 2026Workshop proposal submitted to DAI 2026
TBCNotification from the DAI 2026 chairs
TBCCall for papers opens; challenge environments released
TBCPaper submission deadline
TBCNotifications; challenge entry deadline
29 Nov 2026Workshop — City University of Hong Kong
Jan 2027Workshop report and primitives registry released

DAI 2026 runs 29 November – 2 December 2026 at City University of Hong Kong, under the theme Agentic AI Goes Live: Science, Systems, and Societies. See the conference site for the main program.

People

Organizers

Provisional — the organizing committee is still being confirmed.

Organizer

Botao “Amber” Hu

University of Oxford

Organizer

Helena Rong

New York University Shanghai

Further co-organizers to be announced. The committee is being built to span multi-agent systems, mechanism design, governance and law, and philosophy of computation, with a Hong Kong anchor for regional programme balance.

Invited speakers

To be announced. Speakers will be listed here only once they have confirmed. Invitations that are outstanding will be marked as such rather than presented as commitments.

Program committee

To be announced. The committee will be balanced across multi-agent systems, mechanism design, agent infrastructure, governance and law, and philosophy of computation, and across career stage and region.

Contact

Get in touch

If you work on institutional design, organizational multi-agent systems, agent governance, or the social architecture of agent populations — or if you would like to enter the Institution Design Challenge — we would like to hear from you.

amber@reality.design