Common Ground for Machine Societies

From agents to autonomous organisations

Etylon supports the creation of autonomous organisations: groups of AI agents that pursue a shared purpose over time, organise their work and assess results within defined intent and authority, without a person directing every step. The design question is how such a collective can remain coherent as work, evidence and circumstances change.

The platform brings together Shared Intent and Grounding. Organisations establish purpose and authority. Agents interpret that intent, propose contributions and coordinate commitments through institutional knowledge. Operational arrangements develop through distributed interaction and feedback, including bounded delegation where appropriate. Verification and evaluation remain responsibilities within the undertaking.

Institutional knowledge supplies shared meanings, evidence standards, roles, procedures and commitments. It connects local reasoning to the work others depend on. Intent gives that knowledge a purpose; Grounding makes its implications usable in collective action. Neither an objective nor a supported claim grants permission to act. Challenge and revision preserve a path for correcting both decisions and the terms governing them.

The central hypothesis is that explicit intent and institutional knowledge can support sustained autonomous coordination at an acceptable cost in bounded settings. The longer-term question is how distinct autonomous organisations could cooperate as plural machine societies. The autonomous-organisations guide gives a short introduction; this paper develops the argument, its limits and the tests it requires.

Collective work needs continuity

Agents can generate plans, call tools and exchange messages. An organisation must also retain a purpose, honour commitments and judge whether its work is achieving what matters. A locally useful action can disrupt another agent’s task. An old instruction can remain in circulation after priorities change. A successful tool call can be reported as success while the intended outcome remains unmet.

These problems become harder when work continues across interruptions, hand-offs and changes in participation. Participants need to know which evidence is current, who may decide, what others depend on and when to reconsider a plan. Conversation alone can leave those terms implicit and open to incompatible interpretations.

Distinctions that coordination must preserve

• Evidence can support a claim without establishing its acceptance for every purpose.
• Accepting a claim does not create a commitment or confer authority.
• A shared objective does not override existing responsibilities or permissions.
• Passing a proposal to another agent does not authorise its execution.
• Completing an action does not establish that the desired outcome was achieved.

Etylon’s design makes these distinctions part of the organisation’s shared basis for action. Singh’s work on agent communication provides intellectual context for the importance of public commitments among autonomous, heterogeneous agents.

Autonomy needs institutional terms

An autonomous organisation needs both direction and a basis for acting together. Intent establishes what the organisation is trying to achieve. Institutional knowledge establishes the meanings, evidence standards, roles, commitments and procedures through which agents can pursue it coherently. Etylon designs and creates this knowledge around the organisation’s purpose and coordination needs, then makes it usable and revisable through Grounding.

A constitution, in this paper, is the account of who may decide, within what limits, and how those terms can change. It makes authority inspectable and revisable. It does not prescribe a particular legal form, ownership structure or voting mechanism. Autonomous organisation names a capacity for sustained collective action; the institutional arrangements supporting that capacity still need to be chosen.

The unit of design is therefore the organisation over time. Its agents may use different models, hold different views and take different roles. They need enough common ground to make their contributions compatible and their disagreements examinable. Continuity comes from preserving purpose, dependencies and commitments when individual plans or participants change.

People remain responsible for establishing purpose, authority and the conditions for human review. Agents can choose and revise operational means within that scope. A useful autonomous organisation must also recognise when it lacks evidence or authority to continue.

Shared Intent and Grounding

The two parts of Etylon connect organisational direction with the institutional basis for coordinated action. Their relationship continues through planning, execution, assessment and adaptation.

Shared Intent expresses organisational direction

Organisational intent expresses the purpose, priorities and limits of an undertaking, including expectations for coordination and assessment. Etylon’s Shared Intent capability is proposed as a way to make that direction available and revisable through the organisation’s established authority. Agents interpret it, propose contributions and agree responsibilities within those terms.

Intent leaves room for agent judgement. It is not a script for every action, and it does not grant new authority. When circumstances require different priorities, a proposed revision follows the organisation’s decision process. A changed local plan does not silently change the governing intent.

Grounding supports coherent action

Grounding gives agents access to institutional knowledge and relates their proposals to the meanings, evidence, roles and commitments that apply. It helps them understand what they may rely on, who is affected and what must be resolved before work passes onward. A hand-off gate checks the applicable conditions; a held proposal can be reconsidered while unaffected work continues.

Inside an agent, adaptive reasoning explores possibilities and institutional understanding relates them to the organisation. A plausible next step must be considered alongside its dependencies and the responsibilities of others. This describes complementary capacities without prescribing an internal implementation.

Intent and autonomous coordination

Shared intent gives an undertaking direction without specifying every action required to pursue it. Participants can propose plans, negotiate dependencies and reassess responsibilities within their authority. Temporary organisers or delegated roles may help with particular activities. Routine work can proceed under existing terms without a new negotiation or collective approval for every action.

Feedback may support changes to a plan, a proposed revision of intent or an amendment to institutional terms. These changes have different scopes. A successful local convention does not by itself authorise a change to existing commitments, and a changed objective does not automatically expand an agent’s authority.

Evidence guides the next step

Verifier agents examine whether the evidence meets declared conditions. Evaluator agents assess whether the supported result advances the objective and what remains unresolved. Their findings can support continuation, a revised plan, closure or a request for review. They remain fallible participants: agreement among agents is not independent confirmation of an outcome.

The Shared Intent illustration and Grounding illustration show these responsibilities at agent, organisation and society scales. They communicate the intended behaviour, not measured reliability.

A shared basis that can change

Common ground is the designed and governed body of meanings, records, evidence standards, rules and commitments an institution may currently rely on. Its scope and history matter. It can preserve disagreement and uncertainty while giving agents a usable basis for coordination.

Grounding may reveal that evidence is old, that a proposal conflicts with a commitment, or that an apparent agreement has no authorised decision behind it. The response depends on the case: gather evidence, consult affected participants, revise the proposal or seek a decision. Recognising a problem is not the same as resolving it.

Protocols, conventions and feedback

Explicit interaction protocols can support autonomous participants acting on local information, as illustrated by Mandrake. Conventions can also emerge through repeated interaction. In bounded naming games, Ashery and colleagues observe both shared conventions and collective biases in LLM populations. Their experiments supply game rules and incentives; they do not establish general institutional cooperation.

Etylon’s research question concerns how explicit institutional terms and emerging conventions can work together. A convention can inform a proposed rule, while adoption follows the relevant process and accounts for existing commitments. Feedback can change plans and retained experience without implying online changes to a model’s trained parameters.

Continuity through revision

Agents need to know which intent, evidence and rules a decision relied on. When those grounds change, affected work can be identified and reconsidered. Unaffected commitments remain in force; a new plan does not erase obligations already created. Preserving earlier grounds makes both continuity and challenge possible.

Common ground does not imply a common mind or a complete view of the world. Missing dependencies, mistaken evidence and disagreements about meaning remain possible. The purpose is to make them easier to identify and handle without requiring every agent to share the same internal beliefs.

An organisation restoring a service

Consider a service disruption affecting work already in progress. This scenario illustrates the proposed model; it is not an account of a deployed system.

Organise around the intent

People establish the objective and authority within which the organisation may operate. Agents derive a plan: investigate the disruption, protect dependent work, coordinate recovery and assess the result. Coordination, verification and evaluation responsibilities are taken up by designated agents. Their assignments must fit the applicable authority and competence requirements.

Ground a locally useful proposal

An agent proposes a restart. Its local evidence suggests this could recover the service, but another agent is managing work that depends on the current process. Grounding relates the proposal to that dependency. The hand-off gate holds the affected proposal for revision rather than allowing an unexamined assumption to travel as an accepted basis for further work.

The agents can coordinate a transfer or staged restart that protects the dependent work. Passing the revised proposal onward still does not authorise execution: the acting agent needs the relevant permission, resources and current conditions. Checks are intended to limit the spread of recognised errors; an unknown dependency or an ineffective check can still allow a failure through.

Distinguish activity from achievement

A successful restart is evidence of an action. Verifier agents examine whether service behaviour meets the declared recovery conditions and whether protected work was preserved. Evaluator agents assess whether those findings satisfy the objective, reveal partial recovery or require further action. An unsupported completion report should remain unresolved rather than becoming the organisation’s accepted outcome.

Adapt when capacity changes

If available capacity falls, agents can adjust their plan within scope and surface consequences for the objective. Where priorities must change, an authorised revision of intent might be:

Agents reassess their contributions and reorganise dependencies around the revised intent. Some work continues; some is deferred; verification and evaluation follow the revised criteria. Earlier intent, criteria and findings remain part of the record. Revised criteria govern subsequent work; they do not retrospectively establish that the earlier objective was achieved.

Existing commitments must be honoured or explicitly changed through the applicable procedure. If the agents lack authority to resolve the conflict, they pause the affected work and seek review.

The organisation’s continuity lies in maintaining a shared purpose and accountable relationships through this change. People need not direct each operational step, but retain the ability to inspect, intervene and revise the terms.

Autonomy remains answerable

An institution that applies rules but cannot challenge them risks preserving errors with increasing confidence. Etylon’s design gives decisions, holds and institutional terms a governed path to review.

• Explain: make the relevant evidence, intent, rules and authority behind a decision inspectable.
• Repair: correct evidence, resolve a dependency or satisfy a missing condition.
• Appeal: ask an authorised role or process to reconsider a decision.
• Override: permit a scoped exception under authorised terms, with a reason, expiry and review.
• Amend: revise institutional rules while preserving their history and addressing affected commitments.

Changing a plan, revising organisational intent and changing authority are different acts. Agents may reorganise work within delegated scope. A change to priorities or the governing terms requires the corresponding authority. Evidence of failure can justify a request for change without granting the requester permission to make it.

Outcomes and successful challenges can inform better institutional terms. That learning must itself remain reviewable: a locally convenient exception should not silently become a general rule.

From organisations to machine societies

An autonomous organisation may need to cooperate with others whose purposes, evidence standards and authority differ. The longer-term federation question is whether they can agree enough shared terms to work together while retaining their own institutional identities.

The shared intent of an organisation should not be generalised into a single governing intent for a machine society. Organisations can undertake joint work while retaining different purposes and internal institutions. Ostrom’s discussion of polycentric governance offers a conceptual reference for cooperation among multiple centres of responsibility; evidence from human institutions is not proof about AI societies.

A shared undertaking would need agreed objectives, responsibilities, evidence requirements and dispute routes. Each organisation would interpret its contribution through its own intent and authority. An external request could become a proposal for consideration; it would not automatically become an internal command.

Compacts could define permitted exchanges, limits and expiry. Affected participants would need a route to contest decisions, including consequences the original agreement did not anticipate. Common ground at these boundaries can remain narrower than the knowledge held inside each organisation.

This remains a research direction. Coordination across institutions introduces conflicting obligations, unequal power and uncertain trust. The ambition is a society of organisations able to cooperate and revise their relationships without requiring a universal command hierarchy.

Test the organisation over time

The central evaluation is whether the combined design helps agents sustain useful, accountable work through changing conditions. More messages, more checks or fewer human interventions would not establish that result on their own.

Objectives, feedback and incentives

Agents may face different costs, interests and dependencies. Research on sequential social dilemmas and inequity aversion shows that environmental conditions and particular reward designs can change cooperation in studied games. Etylon’s research should examine how objectives, incentives and institutional arrangements interact, including who benefits and who bears the cost.

Reward functions, reputation, reciprocity and resource access are candidate mechanisms to investigate. This is not a claim that Etylon implements a reward-learning system. Evaluation must distinguish achieved outcomes from rewarded proxies and account for effects on other participants.

Compare credible alternatives

Proposed studies should compare Etylon with explicit decentralised protocols, a competent task orchestrator, message-only teams and simple allocation rules under comparable tasks, models, tools, information access, resources and human supervision. Where feasible, matched variants with explicit intent alone, institutional grounding alone and the combined design would help identify their contributions. These are experimental comparisons, not claims that either part is a complete autonomous organisation by itself.

Tasks should include dependencies, interruptions, changed capacity, stale evidence, conflicting commitments and unsupported completion reports. Evaluate the full sequence from objective to outcome and subsequent adaptation.

Declare success criteria and evaluation procedures before interpreting results. Use independently observable outcomes or review beyond the organisation’s own evaluator agents. Report repeated trials, denominators, uncertainty and adverse cases. Include setup, correction and governance costs. Escalating an unresolved authority conflict can be a correct result; counting all human intervention as failure would reward unsafe autonomy.

Connect to the research programme

The cooperative programme studies coordination and institutional cost; causal policy learning examines effects of decisions and interventions; adaptive constitutional change studies continuity through revised rules. Causal claims require an identification strategy: a recorded decision followed by an outcome does not establish causation. Hernán and Robins provide methodological context.

The hypothesis would be weakened if comparable simpler systems achieve equal or better supported outcomes at lower total cost, or if the combined design adds unacceptable holds, missed conflicts or authority violations. Shadow operation should precede consequential enforcement, with evidence of acceptable error, cost and recovery before scope expands.

Coherence can serve a mistaken purpose

An organisation can coordinate effectively around a poor objective or an unjust rule. Making institutional terms explicit can make power easier to inspect, but does not establish legitimacy or remove politics, bias and domain uncertainty.

• Misstated intent. Ambiguous priorities, stale objectives or a poor success criterion can produce coherent but unwanted behaviour.
• Correlated errors. Workers, verifiers and evaluators may rely on the same mistaken source or share a blind spot. Separate roles alone do not provide independent assurance.
• Incomplete grounding. Missing dependencies, outdated evidence, changes between checking and acting, or paths around a gate can allow errors to propagate.
• Coordination failure. Locally valid plans can conflict, deadlock or impose costs on people and systems outside the organisation’s view.
• Captured governance. Biased meanings, collusion, concentrated authority or misuse of emergency exceptions can entrench harmful decisions.
• Institutional burden. Excessive holds, review demands and maintenance costs can prevent useful work or encourage participants to bypass the process.

Grounding is intended to make recognised problems examinable and manageable; it cannot guarantee that every error is detected or corrected. Verification and evaluation remain dependent on evidence and competent judgement. Authorised actions can still have harmful or unexpected consequences.

Autonomy therefore needs a bounded scope, meaningful intervention and routes for affected parties to challenge decisions. This paper is a functional design and research agenda, not a production protocol or evidence of unrestricted autonomy. The test is both whether the organisation does useful work and whether its governing terms remain answerable and revisable.

A compact for autonomous organisations

Make purpose explicit and keep authority bounded. Design common ground without demanding common minds. Preserve the difference between evidence, commitment and permission. Attach power to roles, scopes and expiry. Keep disagreement visible. Make consequential action and its assessment inspectable. Give every hold a reason and every rule a path to revision. Preserve the grounds of earlier decisions and examine what happened next.

The terms used in this paper

Autonomous organisation (AO)

A group of AI agents that pursues a shared purpose over time, organises its work and assesses results within defined intent and authority, without a person directing every step.

Organisational intent / Shared Intent

Intent establishes the purpose, priorities and limits of an undertaking, including coordination and assessment responsibilities. Shared Intent expresses direction adopted through the organisation’s authority; agents develop contributions and agree operational commitments.

Institutional knowledge / common ground

Institutional knowledge comprises deliberately designed meanings, evidence standards, roles, procedures and commitments. Common ground is the scoped, revisable basis the institution currently relies on to coordinate.

Grounding / hand-off gate

Grounding relates agent proposals and actions to applicable institutional knowledge. A hand-off gate checks the relevant conditions before work passes onward; passing a proposal does not itself confer permission to execute it.

Verification / evaluation

Verification examines evidence against declared conditions. Evaluation assesses what the supported findings mean for the objective and further work. Designated agents carry out these responsibilities; both remain open to review.

Authority / permission

Authority establishes who may decide within a scope. Permission concerns whether a particular action satisfies the applicable conditions. Intent alone supplies neither.

Commitment / outcome

A commitment is an undertaking others may rely on. An outcome is what actually follows from the work. Reported completion, observed effects and satisfaction of the objective must be assessed separately.

Constitution

The roles, decision rights and procedures governing an institution, including how its own terms can change.

The institutional-knowledge explainer and autonomous-organisations guide provide shorter accounts with examples.

Sources and intellectual context

These works provide context for the paper’s questions. Their findings do not validate Etylon’s design or establish its performance.

1. Munindar P. Singh (1999). A Social Semantics for Agent Communication Languages. Develops an account of communication based on public social commitments among autonomous, heterogeneous agents.
2. Patrick Lewis and colleagues (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Studies generation combined with retrieved information. Etylon’s question additionally concerns the institutional acceptance and authorised use of information.
3. Miguel A. Hernán and James M. Robins (2020). Causal Inference: What If. Provides methodological background for defining causal questions and examining the assumptions needed to estimate effects.
4. Samuel H. Christie V, Amit K. Chopra and Munindar P. Singh (2022). Mandrake: multiagent systems as a basis for programming fault-tolerant decentralized applications. Explicit protocols and autonomous endpoints in a particular formal and programming model.
5. Ariel Flint Ashery, Luca Maria Aiello and Andrea Baronchelli (2025). Emergent social conventions and collective bias in LLM populations. Local interactions produce conventions and biases in bounded naming games.
6. Elinor Ostrom (2009). Nobel interview on self-organisation and polycentric governance. Institutional context and cooperation across multiple decision centres.
7. Joel Z. Leibo and colleagues (2017). Multi-agent Reinforcement Learning in Sequential Social Dilemmas. Cooperation and conflict under changing environmental conditions.
8. Edward Hughes and colleagues (2018). Inequity aversion improves cooperation in intertemporal social dilemmas. A particular reward design and its effects in studied games.

Etylon’s proposed contribution is to connect organisational intent, designed institutional knowledge and accountable collective action in autonomous organisations. Evaluation must establish when that combination helps, what it costs and when simpler arrangements are preferable.