What is atomic composability?
Atomic composability is the property that lets one transaction call several applications and either complete every call or undo all of them. It joins two ideas: composability, one application using another's code, assets or output as a part; and atomicity, the database guarantee that a transaction takes effect whole or not at all. Radix's own explainer describes applications combined "as a single transaction" that "either succeeds all together, or fails all together" (Radix Learn: What is Atomic Composability).
The guarantee matters most in DeFi, where a sequence such as swap, deposit and borrow only makes sense complete. Stopped halfway, it leaves the user in a position they never chose. With atomic composability a failed step reverts every step before it, and the user's assets stay as they were, less any fee charged for the failed transaction.
Flash Loans
The standard example is the flash loan: an uncollateralised loan that is borrowed and repaid inside one transaction. The borrower can use the funds for anything in between, but if the repayment step fails, the whole transaction reverts, including the loan itself, so the lender takes no credit risk. The Aave protocol on Ethereum popularised flash loans, and ethereum.org's example is an arbitrage: borrow through a flash loan, buy a token on one exchange, sell it on another, and repay the loan with interest within the same transaction (ethereum.org: Composability). Radix describes the same use, arbitrage "without any credit risk" (Radix Learn).
Which blockchains have atomic composability?
Any blockchain that keeps all its state in one ledger and executes transactions one at a time has it within a single transaction. Ethereum is the best-known case: the arbitrage above is logic that, in ethereum.org's words, "requires combining calls to multiple contracts", and it runs in one transaction. Radix mainnet has the same property today, because one consensus instance settles every transaction against the whole ledger.
The chains differ in where the composition is written. On Ethereum the sequence runs inside a contract that makes each call in turn, either one written for the purpose or a general batching contract. On Radix it is written in the transaction manifest itself: assets returned by one component land on the worktop and pass into the next call, so applications can be combined "in any shape or form" without an integration contract for each combination (Radix Learn). Because the manifest states every withdrawal and deposit, the Radix Wallet can show the user the outcome before they sign (Blind Signing and Transaction Manifests). An AI agent can submit a multi-step strategy the same way, as one manifest that cannot leave it holding half a position.
What are the risks of atomic composability?
A flash loan gives anyone, for the length of one transaction, as much capital as the lender holds, at no risk to the borrower: if the attempt fails, the transaction reverts and they lose only the fee. A flaw that would take a large balance to exploit is therefore open to anyone who finds it. Exploits of DeFi platforms have used flash loans to move spot prices that other contracts then read (Wikipedia: Flash loan).
Radix has its own case. On 27 October 2026 a single manifest borrowed 160,984 ASTRL by flash loan from an Ociswap Flex Pool and, with the loan still open, swapped 7 ASTRL into the same pool for 1.14m $XRD, then repaid the loan. The flash loan supplied the ASTRL the trade needed, and the whole sequence settled in one transaction. Ociswap said the flaw affects only the first version of its Flex Pool and has not published its cause (Ociswap Flex Pool Exploit).
Why do sharding and rollups break atomic composability?
Composing applications in one transaction, with every call committed or none, is synchronous composability. In asynchronous composability each step is its own transaction, so later steps see the results of earlier ones but a failure cannot undo what already committed (Spark: Atomic Composability). Sharding and rollups both push composition from the first kind towards the second.
Both scale a blockchain by splitting its state, and a transaction is only atomic over state that is committed together. On a sharded ledger each shard orders its own transactions, so a transaction that touches accounts on two shards needs both shards to commit it or neither. Ergo's documentation lists the usual tools, from locking and two-phase commit to optimistic execution, against the failures it warns of: assets left locked, or an operation left half done (Ergo Docs: Atomic Composability).
Rollups split Ethereum's activity across separate chains, each with its own consensus, validation and latency, so making a transaction on one rollup final only if its counterpart on another is also final is an unsolved coordination problem. A 2023 paper by Dipankar Sarkar calls the result "fragmented composability" and models what atomic composability across Ethereum rollups would need: buffering, dependency management, concurrency control and zero-knowledge proofs (arXiv:2311.00422).
Radix and Cerberus
Radix's answer is Cerberus, a consensus protocol for multi-shard transactions (Hellings, Hughes, Primero and Sadoghi, 2020). It was specified to keep atomic composability in a sharded network by braiding consensus across only the shards a transaction touches, rather than settling cross-shard operations through an asynchronous bridge.
That remains a specification. Mainnet runs Cerberus unsharded: a single shard group covering the whole ledger, where nothing has to be coordinated across shards. Braiding has never run in production, on Radix or anywhere else, and the Xi’an production candidate does not implement it either: it commits per shard with a two-chain HotStuff-2 protocol, its lead developer having argued that braiding makes shards co-dependent for liveness (hyperscale-rs). Cross-shard atomic composability on Radix is therefore a design target rather than a property of a live network, and any claim that Radix already preserves atomicity "at any scale" is a claim about the whitepaper, not about mainnet.