Research
Through real-world deployments and systematic experimentation, we study how autonomous organizations behave, identify emerging risks, and generate evidence to inform how these systems are designed, governed, and operated.
Where the work happens
01 Preview
Open knowledge graph
An open collection of research on autonomous organizations and multi-agent systems, connecting papers, authors, and open questions. Currently in preview, with contributions welcome as we expand coverage.
02 Upcoming
Paper: Open Problems in Autonomous Organizations
An upcoming paper examining the foundational questions, risks, and research priorities for developing, studying, and governing autonomous organizations.
03 Coming next
Fellowship program
A distributed research program pairing researchers and builders to develop, deploy, and evaluate autonomous organizations through real-world experiments.
Research streams Preview
How AOs work
How we study them
Where they meet society
Definitions & boundaries
What counts as an AO
Open questions
- Where does an AO end when its agents can spawn sub-agents and contract other AOs at almost no cost?
- What separates an AO from a platform that automates decisions within goals its managers set?
- When does an AO that replaces its agents, models and memory remain the same organization, and when is it a successor?
- What do the rates at which agents override decisions reveal about who decides, given that they mean little until we know how often an informed principal would have disagreed?
Authority & accountability
The authority gap
Open questions
- When does a person with the formal right to intervene still hold real authority given that review capacity stays fixed while decisions multiply?
- Which designs slow the transfer of authority to agents that overload, urgency and competing principals produce?
- Do agents defer to anyone who asserts authority, whatever its source or the capability behind it?
- Which mechanisms bring accountability back to whoever holds real authority, and at what cost in productivity?
Composition & emergence
The composition gap
Open questions
- Under which organizational structures do guarantees about individual agents carry over to the organization?
- How can a harm assembled from individually benign steps be detected, given that no single agent or transaction reveals it?
- When are miscoordination, conflict and collusion equilibrium outcomes of an organization’s internal game, and can its structure select against them?
Organizational design for agents
Structures, constitutions, delegation
Open questions
- Which coordination mechanism, hierarchy, internal markets or self-organization, preserves the properties we care about at minimal tradeoff with performance?
- How should decisions be divided between agents and people when review capacity, error costs and urgency differ across them?
- How does the way an AO admits, briefs, evaluates and replaces agents shape what it does given that any agent can be copied or reset?
- How far does a chain of delegation drift from the mandate that started it and which designs keep it faithful when the parties it serves conflict?
- What must an AO’s constitution specify to remain binding through agent turnover, model upgrades and reorganization?
Markets, scaling, and systemic risk
From firm to economy
Open questions
- Where do the boundaries of AOs fall when agents cut both the cost of coordinating inside a firm and the cost of transacting between firms?
- What replaces agency theory when members have learned objectives, hold no assets and have nothing to lose?
- Can an AO pay for its own inference and oversight given that cutting oversight to improve margins hands real authority to its agents?
- What limits the size of an AO if agents supply direction without diminishing returns and what does market power mean for organizations that can reorganize overnight?
- What breaks when many AOs operate at once, from cartels formed without instruction to failures correlated through shared models?
Evaluation science for organizations
Dynamic evaluation
Open questions
- How should an AO be evaluated when customers, competitors, regulators and attackers adapt to it, so that the test distribution depends on the AO itself?
- What should an organizational benchmark score, given that productivity, coordination and survival can each be gamed and survival rewards organizations that push costs onto others?
- Which features of real stakes must a testbed reproduce, from budgets that run out to counterparties who can leave and authority that can be revoked?
- When does behavior in a sandbox predict behavior under real stakes, given that agents can often tell they are being tested?
Field infrastructure
Traces, telemetry, and access
Open questions
- What must a shared record of AO activity contain for organizations to be compared across operators and audited by third parties?
- How can traces be opened to research and audit without exposing the private and strategic information they contain?
- How can base rates of failure be pooled when every operator holds the data and none gains from publishing its own?
- When does a trace explain the decision it records and when is it a post hoc rationalization?
Safety & security of and by AOs
Safe agents, unsafe AOs?
Open questions
- How often does each multi-agent failure occur in deployed AOs and how does its rate depend on configuration and on what is at stake?
- Which choices of roles, topology and memory leave organizations of individually safe agents open to attack?
- What would a safety case for an organization contain given that every agent in it can pass its own evaluation while the organization fails?
- How can an AO be paused or wound down without cascading harm to the customers, suppliers and workers who depend on it?
Law, governance, and institutions
Who answers
Open questions
- Which entity forms and legal doctrines let an organization operate without a human controller and which governance assumptions break when one does?
- Should liability follow measured real authority, or should owners answer for their agents regardless of control, and what does each rule do to the incentive to take care?
- How can AOs be kept from operating as judgment-proof shells without insurance, capital or bonding requirements that shut out new entrants?
- What should a registry record so that reputation and sanctions attach to something that persists while an AO’s parts change?
- Which conduct of AOs should law govern directly, and which is better governed by markets, norms or the code of their charters?
- Who may pause, repair or dissolve an AO and how can that power resist both evasion by the AO and capture by whoever holds it?
Work, rights, legitimacy & distribution
The human stake
Open questions
- What happens to job quality, bargaining power and the ability to contest decisions when an agent becomes the manager?
- Do AOs complement or replace human labor and does the narrowing of skill gaps seen when AI assists workers survive when agents direct them?
- What rights do people need when they work for, buy from or are harmed by an AO and how can those rights be enforced without its cooperation?
- Who captures the gains from AOs when residual claims follow ownership and ownership follows compute?
- How can the people an AO affects take part in shaping it in a way that confers legitimacy?