Agent-assisted representative democracy
Elected bodies remain sovereign. Agents explain bills, compare options, collect instructions and help people challenge decisions.
End Up Here · Research page
Working model / In development
Governance in
a post-agentic
world
When every person has agents acting on their behalf, and advanced systems can produce much more of what people need, what is government for?
Provisional answerGovernment shifts from managing scarcity
to guaranteeing fair access, rights and coordination.
to guaranteeing fair access, rights and coordination.
The system
People, agents
and public institutions.
A plausible system has three layers. The person remains the source of values and rights. A personal agent provides continuous participation: it reads proposals, models consequences, negotiates routine matters and asks its owner when the stakes or uncertainty are high. Public institutions provide common rules, infrastructure, services and remedies.
This is not “AI replaces government”. It is a constitutional democracy with a new civic interface—one that lets people participate continuously without personally processing every question.
Research boundary: this page combines current democratic practice with a speculative systems model. It is not a forecast or claim about present AI capability.
Three probable models
Delegated direct democracy
People vote directly or delegate by topic, time period or confidence. High-stakes questions require stronger human procedures.
Agentic public infrastructure
Government becomes a shared protocol between personal agents: identity, public data, energy, resource access, audits and remedies.
A practical decision cycle
From question
to public action.
1. Frame. An authorised body publishes the question, legal scope, evidence, options, affected groups, decision rule and review date.
2. Deliberate. Agents brief their owners and identify disagreement. For complex questions, randomly selected citizens’ assemblies receive time, evidence and facilitation.
3. Decide. Each person votes, delegates, abstains or requests review. The system records authority, information used and uncertainty.
4. Implement and appeal. Agencies act under law. People can inspect, challenge, withdraw delegation and seek an independent remedy.
What government still has to do
Abundance does not distribute itself.
Even if energy, production and information become abundant, access can remain unequal. Government may need to guarantee civic identity, connectivity, baseline compute, essential services, public-interest models, physical infrastructure and a right to participate without surrendering private life.
It must prevent capture by model providers, infrastructure owners, wealthy institutions, majority coalitions or the state itself. The central question becomes who controls the shared layer between agents—and how that control can be challenged.
Safeguards required
Plural agents
Change providers, compare independent models and keep a portable record of personal instructions.
Human rights floor
Some rights cannot be put to a simple majority vote. Courts, constitutions and independent bodies remain necessary.
Visible uncertainty
Every delegated decision should show authority, evidence, confidence, conflicts and its route to appeal.
Research basis
Current democratic research supports combining representative institutions with direct, participatory and deliberative methods. OECD research documents citizens’ assemblies and participatory budgeting as practical ways to give people informed influence. UNESCO and the Council of Europe emphasise human rights, transparency, oversight, accountability, impact assessment and remedies for AI systems affecting democracy.
OECD participation and deliberation ↗
UNESCO AI ethics recommendation ↗
Council of Europe AI convention ↗
Questions still requiring serious work:
How does an agent represent a changing person? How do we stop coordinated agents manufacturing consent? Which decisions require deliberation rather than aggregation? Who supplies public compute and energy? What happens when agents disagree, fail, are hacked or are captured by their provider?
This is a developing research model. Developing podcast episode ↗