Consensus Mechanisms for Multi-Agent Decisions

When a single agent makes a high-stakes decision, a single point of failure exists. If the agent is compromised, confused, or hallucinating, the wrong action proceeds unchecked. Consensus mechanisms require multiple agents to agree before a decision takes effect, reducing the probability that any single failure causes harm.

Majority Voting

The simplest consensus mechanism is majority voting. Three or more agents independently evaluate the same action. The action proceeds only if a majority approve. This is effective against random errors and single-agent compromise. An attacker must compromise more than half the voting agents to force a malicious action through.

Weighted Voting

Not all agents have equal reliability. A specialized safety agent might have more authority than a general-purpose assistant. Weighted voting assigns each agent a vote weight based on its role, trust level, or historical accuracy. The action proceeds only if the weighted sum of approvals exceeds a threshold.

Quorum-Based Decisions

Quorum mechanisms require a minimum number of agents to participate before any decision is valid. This prevents decisions from proceeding when too few agents are available, which might indicate a denial-of-service attack or system failure.

Integration with Approval Workflows

Authensor's approval workflow system can implement consensus by requiring approvals from multiple agent principals before an action proceeds. Configure the policy to require N-of-M approvals:

rules:
  - action: "payment.send"
    require_approvals: 3
    approval_principals: ["safety-agent", "budget-agent", "compliance-agent", "supervisor-agent"]
    effect: "allow"

Latency Considerations

Consensus adds latency. Every voting agent must evaluate the action before it can proceed. For time-sensitive actions, consider tiered consensus where low-risk actions require fewer votes and high-risk actions require full quorum. Design the tier thresholds in policy so they can be adjusted without code changes.

Consensus is not a replacement for per-agent safety. It is an additional layer that catches failures other mechanisms miss.