Consensus algorithm

Algorand uses a Byzantine agreement protocol that leverages proof of stake.^[18][19][20] As long as a supermajority of the stake is in non-malicious hands, the protocol can tolerate malicious users, achieving consensus without a central authority.

Consensus on Algorand requires three steps to propose, confirm, and write the block to the blockchain.^[21] The steps are proposed soft vote and certified vote.

The first phase (the block proposal phase) uses proof of stake principles. During this phase, a committee of users in the system is selected randomly, though in a manner that is weighted, to propose the new block. The selection of the committee is made via a process called "cryptographic sortition", where each user determines whether they are on the committee by locally executing a Verifiable random function (VRF). If the VRF indicates that the user is chosen, the VRF returns a cryptographic proof that can be used to verify that the user is on the committee. The likelihood that a given user will be on the committee is influenced by the number of ALGO tokens held by that user (the stake).^[17][22][23]

After determining a user is on the block selection committee, that user can build a proposed block and disseminate it to the network for review/analysis during the second phase. The user includes the cryptographic proof from the VRF in their proposed block to demonstrate committee membership.^[22][23]

In the second phase (the block finalization phase), a Byzantine Agreement protocol (called "BA*") is used to vote on the proposed blocks. In this second phase, a new committee is formed via cryptographic sortition. When users have determined that they are in this second-phase voting committee, they analyze the proposed blocks they have received (including verification of first-phase committee membership) and vote on whether any of the blocks should be adopted. If the voting committee achieves consensus on a new block, then the new block is disseminated across the network.^[17][24][25]

Within the Algorand consensus algorithm, membership in both committees changes every time the phase is run. This protects users against targeted attacks, as an attacker will not know in advance which users are going to be on a committee.^[17] Two different Algorand blocks must reach consensus in a different round.^[26][27] According to an external security audit, the model also accounts for timing issues and adversary actions, e.g., when the adversary has control over message delivery.^[28]