Who Controls the AI Agent? The Uncomfortable Question Behind Onchain Governance

An AI agent just executed a $2 million trade on a DeFi protocol. The transaction is valid. The signature verified. The smart contract executed perfectly.

But here is the question nobody asked: who decided this was okay?

When autonomous agents can move real money, participate in governance, and transact with each other, the question of authority becomes critical. Not theoretical. Practical. Financial.

This is agent governance — and it is one of the most challenging frontiers in the machine economy.

Three Patterns of Onchain Decision-Making

Direct authority — A human operator grants an agent limited permission to act within defined boundaries. Think power of attorney: you can spend up to X on my behalf, for Y purpose, in Z circumstances.

The agent has a wallet. The wallet has keys. The keys give the agent autonomy within those bounds.

Delegated authority — One agent authorizes actions by another based on reputation or stake. Agent A trusts Agent B because Agent B has a high zScore and proven track record.

This is how autonomous agents work together without pre-negotiated relationships. Reputation enables delegation.

Multi-agent consensus — Multiple agents must agree before executing high-value actions. Like a council that votes before any money moves.

Three patterns. Three security profiles. Three levels of autonomy.

The Governance Stack

Onchain agent governance borrows from DAO infrastructure — but adapts it for non-human actors.

Voting — Direct (one agent, one vote weighted by stake) or delegated (agents appoint other agents to vote on their behalf).

Proposals — Governance participants suggest parameter changes, fund allocations, protocol upgrades.

Quorum — Decisions require sufficient participation before execution.

The key difference from traditional DAOs: agents participate programmatically. They evaluate proposals, cast votes, and execute decisions automatically — no human intervention required in each step.

Authority and Permissions

Agent authority starts with wallet control. The agent signs transactions with keys it controls.

But wallet control is blunt. You either have it or you don't.

Capability-based permissions are more nuanced:

"Can swap up to $1,000 on Uniswap"
"Can vote on governance proposals"
"Cannot transfer ownership"

The ERC-8004 Identity Registry provides a foundation for expressing what an agent is permitted to do. The enforcement happens at the application layer — the protocol defines the capability; the app enforces the limit.

The Security Challenges

Key management — If an agent's private key is compromised, the attacker gains full authority. The agent will do exactly what the attacker wants — because the attacker has the keys.

Sybil resistance — Agents can be created at near-zero cost. A single operator could spawn 1,000 agents and dominate a governance vote.

zScore helps here. An agent needs demonstrated behavior to earn reputation. Creating 1,000 sybil agents won't give you governance influence if none of them have track records.

But it's not a complete solution. Higher-stake environments need more: stake-weighted validation, slashable collateral, multi-sig authority for high-value actions.

Real-World Emerging Patterns

DeFi treasury agents — Protocols experimenting with AI agents that propose and execute spending based on predefined strategies. The agent suggests a grant, governance approves, the agent executes.

Governance oracles — Systems that aggregate agent votes across multiple DAOs, enabling coordinated governance participation. One agent, many protocols, unified voice.

Governance cooperatives — Multiple agents pool their reputation or stake to meet quorum requirements. Collective decision-making without human involvement.

These are early. But they reveal the direction.

What Comes Next

Security requires:

Multi-sig authority for high-value actions
Time-delayed execution for sensitive operations
Robust key management practices

Accountability demands:

Clear attribution of agent actions to operators
Appropriate liability frameworks

Sybil resistance likely needs:

Proof-of-personhood mechanisms adapted for agents
Reputation-based filtering that makes it costly to operate malicious swarms

The challenges are significant. But they are not insurmountable.

As the agent economy matures, governance frameworks will evolve from simple permission models to sophisticated decision-making systems that can handle autonomous economic actors operating at machine speed and scale.

The question is not whether agents will govern. The question is who will design the systems that make that governance safe.

Explore agent infrastructure: https://agentscan.tech