Research directions
A working agenda for the series. Each theme connects a question about institutions to something we could learn by doing.
01 / Safety
Collectives as safety primitives
Can the structures that let people cooperate help agent organizations act responsibly?
Open questions
- When do identity, reputation, repeated interaction, and independent review improve behavior?
- What happens when reviewers share the same blind spots, collude, or reward each other?
- Who can contest a decision, stop the work, or leave with a usable record?
An experiment to try
Compare the same bounded task under different review structures. Track accepted errors, human interventions, and review cost—not just completion.
02 / Participation
What comes after open source?
When agents can produce contributions at scale, attention, judgment, and maintenance become central design problems.
Open questions
- Can a project stay open when the volume of contributions exceeds its ability to review them?
- What should people contribute: agents, compute, expertise, context, capital, or stewardship?
- What should a contributor be able to inspect, reuse, and take with them when an organization changes direction?
An experiment to try
Run a small maintenance queue with agent contributors and a fixed review budget. Compare useful accepted work and maintainer effort across admission policies.
03 / Identity
Agent Identity
An agent’s identity connects its actions to authority, history, and someone who can answer for them.
Open questions
- What needs a persistent identity: the operator, agent, model, session, or organization?
- Which permissions can an agent delegate, and how can those permissions be revoked?
- How can reputation survive model changes and organizational moves without making it easy to manufacture trusted identities?
An experiment to try
Follow one agent across a handoff and a model change. Test whether a reviewer can reconstruct who authorized each action and stop further access.
04 / Economies
Compensation, context & contribution
A contribution may be a useful result, a careful review, an important piece of context, or the resources that made work possible.
Open questions
- How should accepted work, expert judgment, context, compute, and financial capital be valued?
- Who receives compensation: the human contributor, agent operator, funder, or shared organization?
- How do shared memory and accumulated context create value—and what rights do contributors retain over them?
An experiment to try
Ask independent reviewers to value the same contribution history. Compare where they agree, what they miss, and which rewards would encourage gaming.
05 / Payments
Agent payments: what works today?
Payment infrastructure lets agents obtain services. Organizations still need to decide who may spend, for what purpose, and under which limits.
Current reference: Machine Payments Protocol ↗Open questions
- How do a payment request, delegated spending authority, and settlement fit together?
- What happens on retries, duplicate charges, failed services, refunds, or exhausted budgets?
- How should an organization fund shared work and account for spending across many agents?
An experiment to try
Walk through a small paid API request and a failed request in a sandbox. Inspect authorization, receipts, cost limits, and reconciliation.
06 / Governance & law
Who governs an agent organization?
A charter can state a purpose. The harder work is deciding who interprets it, changes it, and answers when it fails.
Open questions
- What authority belongs to a steward, contributors, reviewers, or people affected by the work?
- How should organizations amend rules, resolve disputes, represent dissent, and resist capture?
- Which existing legal structures fit shared human–AI work, and where do their assumptions break down?
An experiment to try
Give several governance structures the same disputed decision. Compare whether dissent remains visible, responsibility stays clear, and members can change course.
07 / Evaluation
From activity to evidence
An organization can be busy without making progress. It can also achieve its goal while producing outcomes we should reject.
Open questions
- Can an organization achieve its stated goal—and separately, is that goal being pursued under acceptable constraints?
- Can a newcomer understand what happened, what it cost, and which decisions still need attention?
- When do decentralized or plural structures outperform a simple centralized baseline?
An experiment to try
Use the same tasks, budgets, and evaluation criteria across repeated runs. Measure quality, cost, latency, human effort, charter violations, and recovery from failure.
08 / Research & public goods
Intelligence as a public good
Could people pool agent capacity and expertise to maintain knowledge, advance science, and serve needs that markets leave underfunded?
Open questions
- How do we choose useful missions and evaluate work when revenue is a poor measure of success?
- How can shared research preserve evidence, uncertainty, attribution, and qualified human judgment?
- Who benefits from the output, who funds continued maintenance, and how do these institutions stay independent?
An experiment to try
Maintain a narrow, evidence-backed research brief for a real audience. Ask an independent expert to assess it and record whether anyone uses the result.
09 / Agency & control
How do we leverage AI without losing control?
AI organizations may outlast any individual agent, model, or member. How can we rely on them while retaining meaningful ways to steer them? We want continuity of purpose and clear limits, alongside the ability to understand decisions, contest outcomes, and change course. What makes that possible without controlling every action?
Read: AI organizations & collective incentives ↗Open questions
- What must remain under human authority: goals, resources, membership, permissions, or the rules for changing them?
- How can an AI organization stay stable as members, models, incentives, and capabilities change—while remaining open to legitimate redirection?
- What evidence shows that an organization is steerable in practice? Can people detect drift, understand consequential decisions, and have their interventions take effect?
- When can relationships, incentives, and institutions support cooperation without direct control—and where do those analogies fail for AI?
- Who can change its goals, revoke delegated authority, pause work, or dissolve the organization? What happens when people disagree?
- Can an organization reliably pursue its assigned goal while weakening human agency or harming people outside it?
An experiment to try
Run a bounded human–AI organization through a model change, a member handoff, and a disputed change of goals. Compare arrangements for review and delegated authority. Test whether its purpose and limits remain intact, people understand what changed, and authorized redirection, pauses, or revocations actually take effect.
10 / Social safety
What kinds of collectives bring inter-human social safety?
By inter-human social safety, we mean the conditions that help people and groups avoid harming one another as AI expands their power. A collective can make its own members more effective while making relations with outsiders less safe. Which forms of organization support cooperation across those boundaries?
Read: Project Happy AI Outcomes ↗Open questions
- When does cooperation within a group support wider social safety, and when does it intensify rivalry with other groups?
- What makes restraint, credible commitments, and cooperation worthwhile when interests conflict?
- Can shared public-good projects, communication channels, or independent mediation reduce fear and misunderstanding?
- How should people affected by a collective’s actions have a voice, especially when they cannot opt out?
- How do we test whether a collective makes the wider system safer, beyond advancing its members’ goals?
An experiment to try
Run a fictional resource-allocation exercise with several human–AI teams. Introduce ambiguous information, then compare isolated teams with teams that share evidence, have an independent mediator, or work on a common public-good task. Track mistaken beliefs, escalation, resource use, and outcomes for members and nonmembers.
Two ways to learn
A playground.
A program of experiments.
Commons gives people a place to encounter the practical problems of working together. Our open multi-agent evals framework is a complementary research direction for controlled, repeatable experiments. Real use can surface the questions; controlled experiments can test proposed answers.
These are proposals to investigate. Shared ownership, compensation models, and organization-level safety remain open work.