Somewhere on an EVM chain right now, an AI agent is submitting a trade. It came from a workflow, a sequence of tasks: fetch a price feed, check conditions, sign a transaction, and submit. Each step depends on the one before it. That dependency structure has a name. It appears in the routine tools, in the computation graphs that run neural networks, and often at the consensus layer of some of the most actively developed blockchain networks today. It is called a Directed Acyclic Graph, or DAG.
Key Takeaways
- A DAG is a map of dependencies with events pointing forward; nothing loops back.
- It guarantees a consistent order of events without forcing everything through a single queue.
- Developers already use DAGs daily in Git, data pipelines, and AI systems, most just don't call them that.
- The same structure that organizes workflows is now being applied to blockchain consensus.
- DAGs are not a crypto-native idea. It is a technology that distributed networks have borrowed.
When a project promises “thousands of TPS with tiny fees,” the question is whether that comes from real design or clever marketing. Directed Acyclic Graphs (DAGs) are one of the few structural changes that actually let a network process more transactions at once, so understanding them helps you tell the difference.
What is a Directed Acyclic Graph (DAG)?
If you care about fees, confirmation times, or whether a “high-throughput” chain is doing anything new under the hood, you care about DAGs, even if you never plan to touch the code. Some high-throughput networks use DAG-like structures to track which transactions or events do not depend on one another, so more work can be handled in parallel instead of forcing everything into a single line.
A DAG is easiest to picture as dots and arrows. Each dot is an event. Each arrow means “this event had to happen before that.” The arrows only ever point forward, and if you follow them, you never end up back where you started, there are no loops.
- Directed means the arrows go one way: from an earlier event to a later one.
- Acyclic means you cannot make a circle with those arrows. There is no way to start at one event, follow the arrows, and return to it.
- Graph means you are not restricted to a single chain of events. Different paths can branch, join, and exist at the same time.
The point of a DAG is not just that it has no loops, it can represent dependencies in a way that still allows a valid ordering of events.
Once events are drawn this way, there is always at least one sensible way to read them as a timeline: every event appears after the ones it depends on. For a ledger, that provides the structure needed to identify which transactions may be processed independently and which must wait, though the final rules still come from the protocol’s consensus and validity checks.
The concept is not exotic. Git stores commit history as a DAG, where each commit points to one or more parents, branches diverge and merge, and no edge ever points backward to create a loop. The workflow tool, like Apache Airflow is used in most serious data and AI pipelines, defines its workflows literally as DAGs with tasks as nodes and dependencies as edges, and its scheduler uses those dependencies to determine what can run in parallel and what must wait. Feedforward neural network computation graphs are DAGs.
How DAG-Based Ledgers Differ from Blockchains
A standard blockchain enforces a total order on all transactions. Every block has a height, every transaction has a position within that block, and consensus means agreeing on one next block to add to the chain. That constraint keeps double-spend prevention clean - the chain is the single source of truth.
A DAG-based ledger eliminates that constraint. Transactions or events become nodes in a graph, and each new node references one or more earlier ones. The result is a partial order: transactions that genuinely depend on each other have a defined sequence, but two independent transactions do not need a forced ordering between them. Different parts of the network can attach new nodes simultaneously, as long as dependencies are satisfied, so the structure absorbs more parallel activity than a single chain allows.
The throughput gains are real but not automatic. Several first-generation DAG networks paid for higher throughput with centralized node operations or governance. That failure mode drove more recent designs toward using DAG structure only at the consensus layer, not as a replacement for the entire ledger architecture.
DAG in the Wild: Networks That Use It Today
One of the most important shifts in DAG-based crypto over the past years has not been the arrival of new full-DAG ledgers. In order to increase throughput while maintaining a familiar execution environment, blockchain networks have been utilizing DAG more specifically at the consensus layer. That shift matters because the first generation of DAG projects showed how difficult it was to rebuild the whole blockchain stack around a graph.
Sui is the clearest reference point. Its consensus engine, Mysticeti, uses a DAG-based BFT design in which validators can broadcast blocks in parallel instead of waiting for a single leader. Sui’s docs say Mysticeti “only requires 3 rounds of messages to commit blocks from the DAGs”.The graph handles ordering in the background, while developers still build against a conventional execution model.
IOTA was originally built around the Tangle, a DAG-based ledger, and later rebuilt around Sui’s stack, adopting Mysticeti, the MOVE VM, and the parallel execution model. The goal was to give IOTA a more practical route to native smart contracts, faster finality, and a cleaner answer to the long-running problem that had complicated the network’s decentralization story. In 2026, IOTA pushed further with Starfish, a DAG-based BFT protocol designed to improve data availability and resilience under adverse network conditions.
Aptos has moved in the same direction. Its Aptos Labs paper, Narwhal and Tusk: A DAG-based Mempool and Efficient BFT Consensus describes a DAG-based mempool that separates transaction dissemination from ordering. Built on that foundation, the Velociraptr upgrade launched mainnet in late 2025, reducing block time by up to 40% without changing the execution model developers build against.
Hedera belongs in this section as a production example with a different architecture. Its Hashgraph consensus algorithm uses “gossip about gossip” to build a DAG of events and then derives consensus order from that event graph through virtual voting. Hedera presents this design as fast, fair, secure, and asynchronous Byzantine Fault Tolerant, with no mining and no single block producer deciding transaction order.
Sonic is another example of DAG used surgically. Its consensus is DAG-based, and in April 2026 Sonic Labs argued that this design makes post-quantum upgrades easier because changing the signature scheme does not require redesigning the consensus logic itself.
Stable is a live USDT-native Layer 1 blockchain for stablecoins that launched its mainnet in December, 2025. The project says it plans to upgrade to a DAG-based consensus in a future phase, using that structure to improve speed and resilience. The logic is the same as in the networks above: a payments chain designed for large volumes of low-value transfers cannot afford the throughput limits or overload that come with serial block production.
The common thread is that DAG has moved from a full-chain alternative into a more practical tool for solving one narrower problem inside production systems: how to order events quickly without making the whole network wait in a single line.
DAGs, AI Workflows, and On-Chain Agents
The same structure that describes consensus in these networks also describes how modern AI systems organize computation. Neural networks are evaluated over computation graphs: directed, acyclic, and traversed in topological order.
That matters because AI agents also work through dependency chains. An agent does not just “decide” to trade, it often has to fetch data, check conditions, call tools, sign transactions, and confirm settlement in a sequence where each step depends on the one before it.
That structural overlap matters for infrastructure. As AI agents become active participants in on-chain markets, the settlement layer underneath has to handle machine-generated transaction patterns: high volume, low latency, and sensitivity to ordering. By early 2026, AI agent wallets represented a large share of DeFi transaction volume on EVM chains, agents trading, paying for compute, and settling programmatically rather than manually.
If you want to learn more about AI & Web3 Infrastructure read our blog post: AI & Web3 Infrastructure Overlap: Developing the Agent Economy
Research on DAG-based BFT protocols has addressed this workload directly. Designs that separate transaction dissemination from ordering and then derive orders from graph structure after the fact can keep network bandwidth nearly fully utilized under load. Some designs also incorporate explicit ordering-fairness guarantees, making it harder for any node to front-run others by controlling which transactions enter which position first. For agents reacting to the same price signals simultaneously, that property matters more than throughput gains on a sequential chain.
How to Think About DAGs
DAG is an overloaded term. Depending on the design, it may describe the ledger itself, the consensus process, or just the way transactions are ordered. That is why the label alone is not very useful. Two systems can both use DAGs and still make very different trade-offs in speed, finality, decentralization, and security.
The earlier full-ledger models showed how ambitious the idea could be, but they also showed its limits. More recent designs have tended to use DAGs more narrowly, which makes them easier to integrate into existing infrastructure.
The real question is not whether a network uses a DAG. It is what role the DAG plays and what the system gives up to get the benefits it promises.
The information provided by DAIC, including but not limited to research, analysis, data, or other content, is offered solely for informational purposes and does not constitute investment advice, financial advice, trading advice, or any other type of advice. DAIC does not recommend the purchase, sale, or holding of any cryptocurrency or other investment.
